Proofs from the Book [4 ed.] 9783642008559, 9783642008566, 3642008550

Citation preview

Martin Aigner Günter M. Ziegler

Proofs from THE BOOK Fourth Edition

Martin Aigner Günter M. Ziegler

Proofs from THE BOOK Fourth Edition

Including Illustrations by Karl H. Hofmann

123

Martin Aigner FU Berlin Berlin, Germany

ISBN 978-3-642-00855-9 DOI 10.1007/978-3-642-00856-6

Günter M. Ziegler FU Berlin Berlin, Germany

ISBN 978-3-642-00856-6 (eBook)

© Springer-Verlag Berlin Heidelberg 2010, corrected printing 2013 This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Coverdesign: deblik, Berlin Printed on acid-free paper Springer is Part of Springer Science+Business Media www.springer.com

Preface

Paul Erd˝os liked to talk about The Book, in which God maintains the perfect proofs for mathematical theorems, following the dictum of G. H. Hardy that there is no permanent place for ugly mathematics. Erd˝os also said that you need not believe in God but, as a mathematician, you should believe in The Book. A few years ago, we suggested to him to write up a first (and very modest) approximation to The Book. He was enthusiastic about the idea and, characteristically, went to work immediately, filling page after page with his suggestions. Our book was supposed to appear in March 1998 as a present to Erd˝os’ 85th birthday. With Paul’s unfortunate death in the summer of 1996, he is not listed as a co-author. Instead this book is dedicated to his memory. We have no definition or characterization of what constitutes a proof from The Book: all we offer here is the examples that we have selected, hoping that our readers will share our enthusiasm about brilliant ideas, clever insights and wonderful observations. We also hope that our readers will enjoy this despite the imperfections of our exposition. The selection is to a great extent influenced by Paul Erd˝os himself. A large number of the topics were suggested by him, and many of the proofs trace directly back to him, or were initiated by his supreme insight in asking the right question or in making the right conjecture. So to a large extent this book reflects the views of Paul Erd˝os as to what should be considered a proof from The Book. A limiting factor for our selection of topics was that everything in this book is supposed to be accessible to readers whose backgrounds include only a modest amount of technique from undergraduate mathematics. A little linear algebra, some basic analysis and number theory, and a healthy dollop of elementary concepts and reasonings from discrete mathematics should be sufficient to understand and enjoy everything in this book. We are extremely grateful to the many people who helped and supported us with this project — among them the students of a seminar where we discussed a preliminary version, to Benno Artmann, Stephan Brandt, Stefan Felsner, Eli Goodman, Torsten Heldmann, and Hans Mielke. We thank Margrit Barrett, Christian Bressler, Ewgenij Gawrilow, Michael Joswig, Elke Pose, and Jörg Rambau for their technical help in composing this book. We are in great debt to Tom Trotter who read the manuscript from first to last page, to Karl H. Hofmann for his wonderful drawings, and most of all to the late great Paul Erd˝os himself. Berlin, March 1998

Martin Aigner · Günter M. Ziegler

Paul Erd˝os

“The Book”

VI

Preface to the Fourth Edition When we started this project almost fifteen years ago we could not possibly imagine what a wonderful and lasting response our book about The Book would have, with all the warm letters, interesting comments, new editions, and as of now thirteen translations. It is no exaggeration to say that it has become a part of our lives. In addition to numerous improvements, partly suggested by our readers, the present fourth edition contains five new chapters: Two classics, the law of quadratic reciprocity and the fundamental theorem of algebra, two chapters on tiling problems and their intriguing solutions, and a highlight in graph theory, the chromatic number of Kneser graphs. We thank everyone who helped and encouraged us over all these years: For the second edition this included Stephan Brandt, Christian Elsholtz, Jürgen Elstrodt, Daniel Grieser, Roger Heath-Brown, Lee L. Keener, Christian Lebœuf, Hanfried Lenz, Nicolas Puech, John Scholes, Bernulf Weißbach, and many others. The third edition benefitted especially from input by David Bevan, Anders Björner, Dietrich Braess, John Cosgrave, Hubert Kalf, Günter Pickert, Alistair Sinclair, and Herb Wilf. For the present edition, we are particularly grateful to contributions by France Dacar, Oliver Deiser, Anton Dochtermann, Michael Harbeck, Stefan Hougardy, Hendrik W. Lenstra, Günter Rote, Moritz Schmitt, and Carsten Schultz. Moreover, we thank Ruth Allewelt at Springer in Heidelberg as well as Christoph Eyrich, Torsten Heldmann, and Elke Pose in Berlin for their help and support throughout these years. And finally, this book would certainly not look the same without the original design suggested by Karl-Friedrich Koch, and the superb new drawings provided for each edition by Karl H. Hofmann. Berlin, July 2009

Martin Aigner · Günter M. Ziegler

Table of Contents

Number Theory

1

1. Six proofs of the infinity of primes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 2. Bertrand’s postulate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 3. Binomial coefficients are (almost) never powers . . . . . . . . . . . . . . . . . 13 4. Representing numbers as sums of two squares . . . . . . . . . . . . . . . . . . . 17 5. The law of quadratic reciprocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 6. Every finite division ring is a field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 7. Some irrational numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 8. Three times π 2 /6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

Geometry

51

9. Hilbert’s third problem: decomposing polyhedra . . . . . . . . . . . . . . . . . 53 10. Lines in the plane and decompositions of graphs . . . . . . . . . . . . . . . . . 63 11. The slope problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .69 12. Three applications of Euler’s formula . . . . . . . . . . . . . . . . . . . . . . . . . . 75 13. Cauchy’s rigidity theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 14. Touching simplices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 15. Every large point set has an obtuse angle . . . . . . . . . . . . . . . . . . . . . . . 89 16. Borsuk’s conjecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

Analysis

101

17. Sets, functions, and the continuum hypothesis . . . . . . . . . . . . . . . . . . 103 18. In praise of inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 19. The fundamental theorem of algebra . . . . . . . . . . . . . . . . . . . . . . . . . . 127 20. One square and an odd number of triangles . . . . . . . . . . . . . . . . . . . . 131

VIII

Table of Contents 21. A theorem of Pólya on polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 22. On a lemma of Littlewood and Offord . . . . . . . . . . . . . . . . . . . . . . . . . 145 23. Cotangent and the Herglotz trick . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 24. Buffon’s needle problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

Combinatorics

159

25. Pigeon-hole and double counting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 26. Tiling rectangles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 27. Three famous theorems on finite sets . . . . . . . . . . . . . . . . . . . . . . . . . . 179 28. Shuffling cards . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 29. Lattice paths and determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 30. Cayley’s formula for the number of trees . . . . . . . . . . . . . . . . . . . . . . 201 31. Identities versus bijections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207 32. Completing Latin squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213

Graph Theory

219

33. The Dinitz problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .221 34. Five-coloring plane graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 35. How to guard a museum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 36. Turán’s graph theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235 37. Communicating without errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241 38. The chromatic number of Kneser graphs . . . . . . . . . . . . . . . . . . . . . . . 251 39. Of friends and politicians . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257 40. Probability makes counting (sometimes) easy . . . . . . . . . . . . . . . . . . 261

About the Illustrations

270

Index

271

Number Theory

1 Six proofs of the infinity of primes 3 2 Bertrand’s postulate 7 3 Binomial coefficients are (almost) never powers 13 4 Representing numbers as sums of two squares 17 5 The law of quadratic reciprocity 23 6 Every finite division ring is a field 31 7 Some irrational numbers 35 8 Three times π 2 /6 43

“Irrationality and π”

Six proofs of the infinity of primes

Chapter 1

It is only natural that we start these notes with probably the oldest Book Proof, usually attributed to Euclid (Elements IX, 20). It shows that the sequence of primes does not end.  Euclid’s Proof. For any finite set {p1 , . . . , pr } of primes, consider the number n = p1 p2 · · · pr + 1. This n has a prime divisor p. But p is not one of the pi : otherwise p would be a divisor of n and of the product p1 p2 · · · pr , and thus also of the difference n − p1 p2 · · · pr = 1, which is impossible. So a finite set {p1 , . . . , pr } cannot be the collection of all prime numbers.  Before we continue let us fix some notation. N = {1, 2, 3, . . .} is the set of natural numbers, Z = {. . . , −2, −1, 0, 1, 2, . . .} the set of integers, and P = {2, 3, 5, 7, . . .} the set of primes. In the following, we will exhibit various other proofs (out of a much longer list) which we hope the reader will like as much as we do. Although they use different view-points, the following basic idea is common to all of them: The natural numbers grow beyond all bounds, and every natural number n ≥ 2 has a prime divisor. These two facts taken together force P to be infinite. The next proof is due to Christian Goldbach (from a letter to Leonhard Euler 1730), the third proof is apparently folklore, the fourth one is by Euler himself, the fifth proof was proposed by Harry Fürstenberg, while the last proof is due to Paul Erd˝os. n

 Second Proof. Let us first look at the Fermat numbers Fn = 22 +1 for n = 0, 1, 2, . . .. We will show that any two Fermat numbers are relatively prime; hence there must be infinitely many primes. To this end, we verify the recursion n−1  Fk = Fn − 2 (n ≥ 1), k=0

from which our assertion follows immediately. Indeed, if m is a divisor of, say, Fk and Fn (k < n), then m divides 2, and hence m = 1 or 2. But m = 2 is impossible since all Fermat numbers are odd. To prove the recursion we use induction on n. For n = 1 we have F0 = 3 and F1 − 2 = 3. With induction we now conclude n  n−1    Fk = Fk Fn = (Fn − 2)Fn = k=0

k=0 n

n

n+1

= (22 − 1)(22 + 1) = 22

− 1 = Fn+1 − 2.

M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_1, © Springer-Verlag Berlin Heidelberg 2013



F0 = 3 F1 = 5 F2 = 17 F3 = 257 F4 = 65537 F5 = 641 · 6700417 The first few Fermat numbers

4

Six proofs of the infinity of primes

Lagrange’s Theorem If G is a finite (multiplicative) group and U is a subgroup, then |U | divides |G|.  Proof. Consider the binary relation a ∼ b : ⇐⇒ ba−1 ∈ U. It follows from the group axioms that ∼ is an equivalence relation. The equivalence class containing an element a is precisely the coset U a = {xa : x ∈ U }. Since clearly |U a| = |U |, we find that G decomposes into equivalence classes, all of size |U |, and hence that |U | divides |G|.  In the special case when U is a cyclic subgroup {a, a2 , . . . , am } we find that m (the smallest positive integer such that am = 1, called the order of a) divides the size |G| of the group.

 Third Proof. Suppose P is finite and p is the largest prime. We consider the so-called Mersenne number 2p − 1 and show that any prime factor q of 2p − 1 is bigger than p, which will yield the desired conclusion. Let q be a prime dividing 2p − 1, so we have 2p ≡ 1 (mod q). Since p is prime, this means that the element 2 has order p in the multiplicative group Zq \{0} of the field Zq . This group has q − 1 elements. By Lagrange’s theorem (see the box) we know that the order of every element divides the size of the group, that is, we have p | q − 1, and hence p < q.  Now let us look at a proof that uses elementary calculus.  Fourth Proof. Let π(x) := #{p ≤ x : p ∈ P} be the number of primes that are less than or equal to the real number x. We number the primes P = {p1 , p2 , p3 , . . .} in increasing order. Consider the natural logarithm x log x, defined as log x = 1 1t dt.

Now we compare the area below the graph of f (t) = 1t with an upper step function. (See also the appendix on page 10 for this method.) Thus for n ≤ x < n + 1 we have log x ≤ ≤

1 1 1 1 + + ...+ + 2 3 n−1 n  1 , where the sum extends over all m ∈ N which have m only prime divisors p ≤ x.

1+

Since such m can be written in a unique way as a product of the form  kevery p p , we see that the last sum is equal to p≤x

  1  . pk

p∈P p≤x

1

k≥0

The inner sum is a geometric series with ratio p1 , hence log x ≤ 1

2

p∈P p≤x

n n+1

Steps above the function f (t) =

1 t



1 1−

1 p

=

 p∈P p≤x

π(x)  pk p = . p−1 pk − 1 k=1

Now clearly pk ≥ k + 1, and thus 1 1 k+1 pk = 1+ ≤ 1+ = , pk − 1 pk − 1 k k and therefore

 k+1 = π(x) + 1. k

π(x)

log x ≤

k=1

Everybody knows that log x is not bounded, so we conclude that π(x) is unbounded as well, and so there are infinitely many primes. 

Six proofs of the infinity of primes

5

 Fifth Proof. After analysis it’s topology now! Consider the following curious topology on the set Z of integers. For a, b ∈ Z, b > 0, we set Na,b = {a + nb : n ∈ Z}. Each set Na,b is a two-way infinite arithmetic progression. Now call a set O ⊆ Z open if either O is empty, or if to every a ∈ O there exists some b > 0 with Na,b ⊆ O. Clearly, the union of open sets is open again. If O1 , O2 are open, and a ∈ O1 ∩ O2 with Na,b1 ⊆ O1 and Na,b2 ⊆ O2 , then a ∈ Na,b1 b2 ⊆ O1 ∩ O2 . So we conclude that any finite intersection of open sets is again open. So, this family of open sets induces a bona fide topology on Z. Let us note two facts: (A) Any nonempty open set is infinite. (B) Any set Na,b is closed as well. Indeed, the first fact follows from the definition. For the second we observe Na,b = Z \

b−1 

Na+i,b ,

i=1

which proves that Na,b is the complement of an open set and hence closed. So far the primes have not yet entered the picture — but here they come. Since any number n = 1, −1 has a prime divisor p, and hence is contained in N0,p , we conclude  Z \ {1, −1} = N0,p . p∈P

Now if P were finite, then p∈P N0,p would be a finite union of closed sets (by (B)), and hence closed. Consequently, {1, −1} would be an open set, in violation of (A).   Sixth Proof. Our final proof goes a considerable step further and demonstrates

not only that there are infinitely many primes, but also that the series p∈P 1p diverges. The first proof of this important result was given by Euler (and is interesting in its own right), but our proof, devised by Erd˝os, is of compelling beauty. Let p1 , p2 , p

3 , . . . be the sequence of primes in increasing order, and assume that p∈P p1 converges. Then there must be a natural number k

such that i≥k+1 p1i < 12 . Let us call p1 , . . . , pk the small primes, and pk+1 , pk+2 , . . . the big primes. For an arbitrary natural number N we therefore find  N N . (1) < pi 2 i≥k+1

“Pitching flat rocks, infinitely”

6

Six proofs of the infinity of primes Let Nb be the number of positive integers n ≤ N which are divisible by at least one big prime, and Ns the number of positive integers n ≤ N which have only small prime divisors. We are going to show that for a suitable N Nb + Ns < N, which will be our desired contradiction, since by definition Nb + Ns would have to be equal to N . To estimate Nb note that pNi counts the positive integers n ≤ N which are multiples of pi . Hence by (1) we obtain Nb ≤

 N N . < pi 2

(2)

i≥k+1

Let us now look at Ns . We write every n ≤ N which has only small prime divisors in the form n = an b2n , where an is the square-free part. Every an is thus a product of different small primes, and we conclude that √ there√are precisely 2k different square-free parts. Furthermore, as b ≤ n ≤ N, n √ we find that there are at most N different square parts, and so √ Ns ≤ 2 k N . √ Since (2) holds for any N , it remains to find a number N with 2k N ≤ N2 √ or 2k+1 ≤ N , and for this N = 22k+2 will do. 

References [1] B. A RTMANN : Euclid — The Creation of Mathematics, Springer-Verlag, New York 1999. 1 ˝ : Über die Reihe , Mathematica, Zutphen B 7 (1938), 1-2. [2] P. E RD OS p [3] L. E ULER : Introductio in Analysin Infinitorum, Tomus Primus, Lausanne 1748; Opera Omnia, Ser. 1, Vol. 8. [4] H. F ÜRSTENBERG : On the infinitude of primes, Amer. Math. Monthly 62 (1955), 353.

Bertrand’s postulate

Chapter 2

We have seen that the sequence of prime numbers 2, 3, 5, 7, . . . is infinite. To see that the size of its gaps is not bounded, let N := 2 · 3 · 5 · · · p denote the product of all prime numbers that are smaller than k + 2, and note that none of the k numbers N + 2, N + 3, N + 4, . . . , N + k, N + (k + 1) is prime, since for 2 ≤ i ≤ k + 1 we know that i has a prime factor that is smaller than k + 2, and this factor also divides N , and hence also N + i. With this recipe, we find, for example, for k = 10 that none of the ten numbers 2312, 2313, 2314, . . ., 2321 is prime. But there are also upper bounds for the gaps in the sequence of prime numbers. A famous bound states that “the gap to the next prime cannot be larger than the number we start our search at.” This is known as Bertrand’s postulate, since it was conjectured and verified empirically for n < 3 000 000 by Joseph Bertrand. It was first proved for all n by Pafnuty Chebyshev in 1850. A much simpler proof was given by the Indian genius Ramanujan. Our Book Proof is by Paul Erd˝os: it is taken from Erd˝os’ first published paper, which appeared in 1932, when Erd˝os was 19. Bertrand’s postulate. For every n ≥ 1, there is some prime number p with n < p ≤ 2n.

  Proof. We will estimate the size of the binomial coefficient 2n n carefully enough to see that if it didn’t have any prime factors in the range n < p ≤ 2n, then it would be “too small.” Our argument is in five steps. (1) We first prove Bertrand’s postulate for n < 4000. For this one does not need to check 4000 cases: it suffices (this is “Landau’s trick”) to check that 2, 3, 5, 7, 13, 23, 43, 83, 163, 317, 631, 1259, 2503, 4001 is a sequence of prime numbers, where each is smaller than twice the previous one. Hence every interval {y : n < y ≤ 2n}, with n ≤ 4000, contains one of these 14 primes. M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_2, © Springer-Verlag Berlin Heidelberg 2013

Joseph Bertrand

8

Bertrand’s postulate (2) Next we prove that  p ≤ 4x−1

for all real x ≥ 2,

(1)

p≤x

where our notation — here and in the following — is meant to imply that the product is taken over all prime numbers p ≤ x. The proof that we present for this fact uses induction on the number of these primes. It is not from Erd˝os’ original paper, but it is also due to Erd˝os (see the margin), and it is a true Book Proof. First we note that if q is the largest prime with q ≤ x, then   p = p and 4q−1 ≤ 4x−1 . p≤x

p≤q

Thus it suffices to check (1) for the case where x = q is a prime number. For q = 2 we get “2 ≤ 4,” so we proceed to consider odd primes q = 2m + 1. (Here we may assume, by induction, that (1) is valid for all integers x in the set {2, 3, . . . , 2m}.) For q = 2m + 1 we split the product and compute      m 2m + 1 ≤ 4m 22m = 42m. p = p· p ≤ 4 m p≤2m+1

p≤m+1

m+1 2n appear at most once in 2n n . Furthermore — and this, according to Erd˝os, is the key  for his proof

fact — primes p that satisfy 23 n < p ≤ n do not divide 2n n at all! Indeed, 3p > 2n implies (for n ≥ 3, and hence p ≥ 3) that p and 2p are the only multiples of p that appear as factors in the numerator of (2n)! n!n! , while we get two p-factors in the denominator.

 (4) Now we are ready to estimate 2n n . For n ≥ 3, using an estimate from page 12 for the lower bound, we get      4n 2n ≤ ≤ 2n · p · p 2n n √ √ 2 p≤ 2n

2n n = 1. 1 = n0 < n1 < · · · < n/2 Here x resp. x denotes the number x rounded down resp. rounded up to the nearest integer. From the asymptotic formulas for the factorials mentioned above one can obtain very precise estimates for the sizes of binomial coefficients. However, we will only need  very weak and simple estimates in this book, such as the following: nk ≤ 2n for all k, while for n ≥ 2 we have   n 2n ≥ , n/2

n with equality only for n = 2. In particular, for n ≥ 1,   2n 4n ≥ . n 2n

n  This holds since n/2 , a middle binomial coefficient, is the largest entry

n n n n

n  in the sequence 0 + n , 1 , 2 , . . . , n−1 , whose sum is 2n , and whose 2n average is thus n . On the other hand, we note the upper bound for binomial coefficients   n n(n − 1) · · · (n − k + 1) nk nk = ≤ ≤ k−1 , k k! k! 2 which is a reasonably good estimate for the “small” binomial coefficients at the tails of the sequence, when n is large (compared to k).

References ˝ : Beweis eines Satzes von Tschebyschef, Acta Sci. Math. (Szeged) 5 [1] P. E RD OS (1930-32), 194-198.

[2] R. L. G RAHAM , D. E. K NUTH & O. PATASHNIK : Concrete Mathematics. A Foundation for Computer Science, Addison-Wesley, Reading MA 1989. [3] G. H. H ARDY & E. M. W RIGHT: An Introduction to the Theory of Numbers, fifth edition, Oxford University Press 1979. [4] P. R IBENBOIM : The New Book of Prime Number Records, Springer-Verlag, New York 1989.

Binomial coefficients are (almost) never powers

Chapter 3

There is an epilogue to Bertrand’s postulate which leads to a beautiful result on binomial coefficients. In 1892 Sylvester strengthened Bertrand’s postulate in the following way: If n ≥ 2k, then at least one of the numbers n, n − 1, . . . , n − k + 1 has a prime divisor p greater than k. Note that for n = 2k we obtain precisely Bertrand’s postulate. In 1934, Erd˝os gave a short and elementary Book Proof of Sylvester’s result, running along the lines of his proof of Bertrand’s postulate. There is an equivalent way of stating Sylvester’s theorem: The binomial coefficient   n n(n − 1) · · · (n − k + 1) = k k!

(n ≥ 2k)

always has a prime factor p > k. With this observation in mind, we turn to another one of Erd˝os’ jewels.  When is nk equal to a power m ? It is easy to see that there  are infinitely many solutions for k =  = 2, that is, of the equation n2 = m2 . Indeed,

 2 . To see this, set n(n − 1) = 2m2 . if n2 is a square, then so is (2n−1) 2 It follows that (2n − 1)2 ((2n − 1)2 − 1) = (2n − 1)2 4n(n − 1) = 2(2m(2n − 1))2 , and hence

  (2n − 1)2 = (2m(2n − 1))2 . 2

 Beginning with 92 = 62 we thus obtain infinitely many solutions — the

 next one is 289 = 2042. However, this does not yield all solutions. For 

502

= 11892. For example, 2 = 352 starts another series, as does 1682 2  n 2 k = 3 it is known that 3 = m has the unique solution n = 50, m = 140. But now we are at the end of the line. For k ≥ 4 and any  ≥ 2 no solutions exist, and this is what Erd˝os proved by an ingenious argument.

 Theorem. The equation nk = m has no integer solutions with  ≥ 2 and 4 ≤ k ≤ n − 4.

M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_3, © Springer-Verlag Berlin Heidelberg 2013

50 3

= 1402

is the only solution for k = 3,  = 2

14

Binomial coefficients are (almost) never powers

 n  .  Proof. Note first that we may assume n ≥ 2k because of nk = n−k

n  Suppose the theorem is false, and that k = m . The proof, by contradiction, proceeds in the following four steps.

 (1) By Sylvester’s theorem, there is a prime factor p of nk greater than k, hence p divides n(n − 1) · · · (n − k + 1). Clearly, only one of the factors n − i can be a multiple of p (because of p > k), and we conclude p | n − i, and therefore n ≥ p > k  ≥ k 2 . (2) Consider any factor n − j of the numerator and write it in the form n − j = aj mj , where aj is not divisible by any nontrivial -th power. We note by (1) that aj has only prime divisors less than or equal to k. We want to show next that ai = aj for i = j. Assume to the contrary that ai = aj for some i < j. Then mi ≥ mj + 1 and k

>

(n − i) − (n − j) = aj (mi − mj ) ≥ aj ((mj + 1) − mj )

>

≥ (aj mj )1/2 ≥ (n − k + 1)1/2 aj m−1 j

≥

( n2 + 1)1/2 > n1/2 ,

which contradicts n > k 2 from above. (3) Next we prove that the ai ’s are the integers 1, 2, . . . , k in some order. (According to Erd˝os, this is the crux of the proof.) Since we already know that they are all distinct, it suffices to prove that a0 a1 · · · ak−1 divides k!.

 Substituting n − j = aj mj into the equation nk = m , we obtain a0 a1 · · · ak−1 (m0 m1 · · · mk−1 ) = k!m . Cancelling the common factors of m0 · · · mk−1 and m yields a0 a1 · · · ak−1 u = k!v  with gcd(u, v) = 1. It remains to show that v = 1. If not, then v contains a prime divisor p. Since gcd(u, v) = 1, p must be a prime divisor of a0 a1 · · · ak−1 and hence is less than or equal to k. By the

theorem of Legendre (see page 8) we know that k! contains p to the power i≥1 pki . We now estimate the exponent of p in n(n − 1) · · · (n − k + 1). Let i be a positive integer, and let b1 < b2 < · · · < bs be the multiples of pi among n, n − 1, . . . , n − k + 1. Then bs = b1 + (s − 1)pi and hence (s − 1)pi = bs − b1 ≤ n − (n − k + 1) = k − 1, which implies s ≤

k − 1 pi

+1 ≤

k pi

+ 1.

Binomial coefficients are (almost) never powers

15

So for each i the number of multiples of pi among n, . . . , n−k+1, and hence among the aj ’s, is bounded by pki + 1. This implies that the exponent of p in a0 a1 · · · ak−1 is at most −1   k i=1

pi

 +1

with the reasoning that we used for Legendre’s theorem in Chapter 2. The only difference is that this time the sum stops at i =  − 1, since the aj ’s contain no -th powers. Taking both counts together, we find that the exponent of p in v  is at most −1   k i=1

pi

  k +1 − ≤  − 1, pi i≥1

and we have our desired contradiction, since v  is an -th power. This suffices already to settle the case  = 2. Indeed, since k ≥ 4 one of the ai ’s must be equal to 4, but the ai ’s contain no squares. So let us now assume that  ≥ 3. (4) Since k ≥ 4, we must have ai1 = 1, ai2 = 2, ai3 = 4 for some i1 , i2 , i3 , that is, n − i1 = m1 , n − i2 = 2m2 , n − i3 = 4m3 . We claim that (n − i2 )2 = (n − i1 )(n − i3 ). If not, put b = n − i2 and n − i1 = b − x, n − i3 = b + y, where 0 < |x|, |y| < k. Hence b2 = (b − x)(b + y)

or

(y − x)b = xy,

where x = y is plainly impossible. Now we have by part (1) |xy| = b|y − x| ≥ b > n − k > (k − 1)2 ≥ |xy|, which is absurd. So we have m22 = m1 m3 , where we assume m22 > m1 m3 (the other case being analogous), and proceed to our last chains of inequalities. We obtain 2(k − 1)n

>

n2 − (n − k + 1)2 > (n − i2 )2 − (n − i1 )(n − i3 )

=

   4[m2 2 − (m1 m3 ) ] ≥ 4[(m1 m3 + 1) − (m1 m3 ) ]

≥

−1 4m−1 1 m3 .

Since  ≥ 3 and n > k  ≥ k 3 > 6k, this yields 2(k − 1)nm1 m3

> >

4m1 m3 = (n − i1 )(n − i3 ) n (n − k + 1)2 > 3(n − )2 > 2n2 . 6

We see that our analysis so far agrees  = 1402 , as with 50 3 50 = 2 · 52 49 = 1 · 72 48 = 3 · 42 and 5 · 7 · 4 = 140.

16

Binomial coefficients are (almost) never powers Now since mi ≤ n1/ ≤ n1/3 we finally obtain kn2/3 ≥ km1 m3 > (k − 1)m1 m3 > n, or k 3 > n. With this contradiction, the proof is complete.



References ˝ : A theorem of Sylvester and Schur, J. London Math. Soc. 9 (1934), [1] P. E RD OS 282-288. ˝ : On a diophantine equation, J. London Math. Soc. 26 (1951), [2] P. E RD OS 176-178.

[3] J. J. S YLVESTER : On arithmetical series, Messenger of Math. 21 (1892), 1-19, 87-120; Collected Mathematical Papers Vol. 4, 1912, 687-731.

Representing numbers as sums of two squares

Which numbers can be written as sums of two squares? This question is as old as number theory, and its solution is a classic in the field. The “hard” part of the solution is to see that every prime number of the form 4m + 1 is a sum of two squares. G. H. Hardy writes that this two square theorem of Fermat “is ranked, very justly, as one of the finest in arithmetic.” Nevertheless, one of our Book Proofs below is quite recent. Let’s start with some “warm-ups.” First, we need to distinguish between the prime p = 2, the primes of the form p = 4m + 1, and the primes of the form p = 4m + 3. Every prime number belongs to exactly one of these three classes. At this point we may note (using a method “à la Euclid”) that there are infinitely many primes of the form 4m + 3. In fact, if there were only finitely many, then we could take pk to be the largest prime of this form. Setting Nk := 22 · 3 · 5 · · · pk − 1

Chapter 4

1 2 3 4 5 6 7 8 9 10 11

= = = = = = = = = = = .. .

12 + 02 12 + 12 22 + 02 22 + 12

22 + 22 32 + 32 +

(where p1 = 2, p2 = 3, p3 = 5, . . . denotes the sequence of all primes), we find that Nk is congruent to 3 (mod 4), so it must have a prime factor of the form 4m + 3, and this prime factor is larger than pk — contradiction. Our first lemma characterizes the primes for which −1 is a square in the field Zp (which is reviewed in the box on the next page). It will also give us a quick way to derive that there are infinitely many primes of the form 4m + 1. Lemma 1. For primes p = 4m + 1 the equation s2 ≡ −1 (mod p) has two solutions s ∈ {1, 2, . . ., p−1}, for p = 2 there is one such solution, while for primes of the form p = 4m + 3 there is no solution.  Proof. For p = 2 take s = 1. For odd p, we construct the equivalence relation on {1, 2, . . . , p − 1} that is generated by identifying every element with its additive inverse and with its multiplicative inverse in Zp . Thus the “general” equivalence classes will contain four elements {x, −x, x, −x} since such a 4-element set contains both inverses for all its elements. However, there are smaller equivalence classes if some of the four numbers are not distinct: M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_4, © Springer-Verlag Berlin Heidelberg 2013

Pierre de Fermat

18

Representing numbers as sums of two squares • x ≡ −x is impossible for odd p. • x ≡ x is equivalent to x2 ≡ 1. This has two solutions, namely x = 1 and x = p − 1, leading to the equivalence class {1, p − 1} of size 2. • x ≡ −x is equivalent to x2 ≡ −1. This equation may have no solution or two distinct solutions x0 , p − x0 : in this case the equivalence class is {x0 , p − x0 }.

For p = 11 the partition is {1, 10}, {2, 9, 6, 5}, {3, 8, 4, 7}; for p = 13 it is {1, 12}, {2, 11, 7, 6}, {3, 10, 9, 4}, {5, 8}: the pair {5, 8} yields the two solutions of s2 ≡ −1 mod 13.

The set {1, 2, . . . , p − 1} has p − 1 elements, and we have partitioned it into quadruples (equivalence classes of size 4), plus one or two pairs (equivalence classes of size 2). For p − 1 = 4m + 2 we find that there is only the one pair {1, p− 1}, the rest is quadruples, and thus s2 ≡ −1 (mod p) has no solution. For p − 1 = 4m there has to be the second pair, and this contains the two solutions of s2 ≡ −1 that we were looking for.  Lemma 1 says that every odd prime dividing a number M 2 + 1 must be of the form 4m + 1. This implies that there are infinitely many primes of this form: Otherwise, look at (2 · 3 · 5 · · · qk )2 + 1, where qk is the largest such prime. The same reasoning as above yields a contradiction.

Prime fields If p is a prime, then the set Zp = {0, 1, . . . , p − 1} with addition and multiplication defined “modulo p” forms a finite field. We will need the following simple properties: + 0 1 2 3 4

0 0 1 2 3 4

1 1 2 3 4 0

2 2 3 4 0 1

3 3 4 0 1 2

4 4 0 1 2 3

· 0 1 2 3 4

0 0 0 0 0 0

1 0 1 2 3 4

2 0 2 4 1 3

3 0 3 1 4 2

4 0 4 3 2 1

Addition and multiplication in Z5

• For x ∈ Zp , x = 0, the additive inverse (for which we usually write −x) is given by p − x ∈ {1, 2, . . . , p − 1}. If p > 2, then x and −x are different elements of Zp . • Each x ∈ Zp \{0} has a unique multiplicative inverse x ∈ Zp \{0}, with xx ≡ 1 (mod p). The definition of primes implies that the map Zp → Zp , z → xz is injective for x = 0. Thus on the finite set Zp \{0} it must be surjective as well, and hence for each x there is a unique x = 0 with xx ≡ 1 (mod p). • The squares 02 , 12 , 22 , . . . , h2 define different elements of Zp , for h = p2 . This is since x2 ≡ y 2 , or (x + y)(x − y) ≡ 0, implies that x ≡ y or that x ≡ −y. The 1 + p2 elements 02 , 12 , . . . , h2 are called the squares in Zp . At this point, let us note “on the fly” that for all primes there are solutions for x2 + y 2 ≡ −1 (mod p). In fact, there are p2 + 1 distinct squares x2 in Zp , and there are p2 + 1 distinct numbers of the form −(1 + y 2 ). These two sets of numbers are too large to be disjoint, since Zp has only p elements, and thus there must exist x and y with x2 ≡ −(1 + y 2 ) (mod p).

Representing numbers as sums of two squares

19

Lemma 2. No number n = 4m + 3 is a sum of two squares.  Proof. The square of any even number is (2k)2 = 4k 2 ≡ 0 (mod 4), while squares of odd numbers yield (2k+1)2 = 4(k 2 +k)+1 ≡ 1 (mod 4). Thus any sum of two squares is congruent to 0, 1 or 2 (mod 4).  This is enough evidence for us that the primes p = 4m + 3 are “bad.” Thus, we proceed with “good” properties for primes of the form p = 4m + 1. On the way to the main theorem, the following is the key step. Proposition. Every prime of the form p = 4m + 1 is a sum of two squares, that is, it can be written as p = x2 + y 2 for some natural numbers x, y ∈ N. We shall present here two proofs of this result — both of them elegant and surprising. The first proof features a striking application of the “pigeonhole principle” (which we have already used “on the fly” before Lemma 2; see Chapter 25 for more), as well as a clever move to arguments “modulo p” and back. The idea is due to the Norwegian number theorist Axel Thue. √  Proof. Consider the pairs (x , y ) of integers with 0 ≤ x , y ≤ p, that √ √ is, x , y ∈ {0, 1, . . . , p }. There are ( p + 1)2 such pairs. Using the √ estimate x + 1 > x for x = p, we see that we have more than p such pairs of integers. Thus for any s ∈ Z, it is impossible that all the values x − sy produced by the pairs (x , y ) are distinct modulo p. That is, for every s there are two distinct pairs √ (x , y ), (x , y ) ∈ {0, 1, . . . , p }2 with x − sy ≡ x − sy (mod p). Now we take differences: We have x − x ≡ s(y − y ) (mod p). Thus if we define x := |x − x |, y := |y − y |, then we get √ (x, y) ∈ {0, 1, . . . , p }2 with x ≡ ±sy (mod p). Also we know that not both x and y can be zero, because the pairs (x , y ) and (x , y ) are distinct. Now let s be a solution of s2 ≡ −1 (mod p), which exists by Lemma 1. Then x2 ≡ s2 y 2 ≡ −y 2 (mod p), and so we have produced (x, y) ∈ Z2

with

0 < x2 + y 2 < 2p

and

x2 + y 2 ≡ 0 (mod p).

But p is the only number between 0 and 2p that is divisible by p. Thus x2 + y 2 = p: done!  Our second proof for the proposition — also clearly a Book Proof — was discovered by Roger Heath-Brown in 1971 and appeared in 1984. (A condensed “one-sentence version” was given by Don Zagier.) It is so elementary that we don’t even need to use Lemma 1. Heath-Brown’s argument features three linear involutions: a quite obvious one, a hidden one, and a trivial one that gives “the final blow.” The second, unexpected, involution corresponds to some hidden structure on the set of integral solutions of the equation 4xy + z 2 = p.

√ For p = 13, p = 3 we consider x , y  ∈ {0, 1, 2, 3}. For s = 5, the sum x −sy  (mod 13) assumes the following values:  Qy 0 1 2 3 x Q 0 0 8 3 11 1 1 9 4 12 2 10 5 0 2 3 3 11 6 1

20

Representing numbers as sums of two squares  Proof. We study the set S := {(x, y, z) ∈ Z3 : 4xy + z 2 = p,

x > 0,

y > 0}.

This set is finite. Indeed, x ≥ 1 and y ≥ 1 implies y ≤ p4 and x ≤ p4 . So there are only finitely many possible values for x and y, and given x and y, there are at most two values for z. 1. The first linear involution is given by f : S −→ S,

(x, y, z) −→ (y, x, −z),

that is, “interchange x and y, and negate z.” This clearly maps S to itself, and it is an involution: Applied twice, it yields the identity. Also, f has no fixed points, since z = 0 would imply p = 4xy, which is impossible. Furthermore, f maps the solutions in T := {(x, y, z) ∈ S : z > 0} to the solutions in S\T , which satisfy z < 0. Also, f reverses the signs of x − y and of z, so it maps the solutions in

T

U := {(x, y, z) ∈ S : (x − y) + z > 0}

f

to the solutions in S\U . For this we have to see that there is no solution with (x−y)+z = 0, but there is none since this would give p = 4xy+z 2 = 4xy + (x − y)2 = (x + y)2 . What do we get from the study of f ? The main observation is that since f maps the sets T and U to their complements, it also interchanges the elements in T \U with these in U \T . That is, there is the same number of solutions in U that are not in T as there are solutions in T that are not in U — so T and U have the same cardinality.

U

2. The second involution that we study is an involution on the set U : g : U −→ U,

g

(x, y, z) −→ (x − y + z, y, 2y − z).

First we check that indeed this is a well-defined map: If (x, y, z) ∈ U , then x − y + z > 0, y > 0 and 4(x − y + z)y + (2y − z)2 = 4xy + z 2 , so g(x, y, z) ∈ S. By (x − y + z) − y + (2y − z) = x > 0 we find that indeed g(x, y, z) ∈ U . Also g is an involution: g(x, y, z) = (x − y + z, y, 2y − z) is mapped by g to ((x − y + z) − y + (2y − z), y, 2y − (2y − z)) = (x, y, z). And finally: g has exactly one fixed point: (x, y, z) = g(x, y, z) = (x − y + z, y, 2y − z)

U

holds exactly if y = z: But then p = 4xy + y 2 = (4x + y)y, which holds only for y = 1 = z, and x = p−1 4 . But if g is an involution on U that has exactly one fixed point, then the cardinality of U is odd.

Representing numbers as sums of two squares

21

3. The third, trivial, involution that we study is the involution on T that interchanges x and y: h : T −→ T,

(x, y, z) −→ (y, x, z).

This map is clearly well-defined, and an involution. We combine now our knowledge derived from the other two involutions: The cardinality of T is equal to the cardinality of U , which is odd. But if h is an involution on a finite set of odd cardinality, then it has a fixed point: There is a point (x, y, z) ∈ T with x = y, that is, a solution of p = 4x2 + z 2 = (2x)2 + z 2 .



Note that this proof yields more — the number of representations of p in the form p = x2 + (2y)2 is odd for all primes of the form p = 4m + 1. (The representation is actually unique, see [3].) Also note that both proofs are not effective: Try to find x and y for a ten digit prime! Efficient ways to find such representations as sums of two squares are discussed in [1] and [7]. The following theorem completely answers the question which started this chapter. Theorem. A natural number n can be represented as a sum of two squares if and only if every prime factor of the form p = 4m + 3 appears with an even exponent in the prime decomposition of n.  Proof. Call a number n representable if it is a sum of two squares, that is, if n = x2 + y 2 for some x, y ∈ N0 . The theorem is a consequence of the following five facts. (1) 1 = 12 + 02 and 2 = 12 + 12 are representable. Every prime of the form p = 4m + 1 is representable. (2) The product of any two representable numbers n1 = x21 + y12 and n2 = x22 + y22 is representable: n1 n2 = (x1 x2 + y1 y2 )2 + (x1 y2 − x2 y1 )2 . (3) If n is representable, n = x2 + y 2 , then also nz 2 is representable, by nz 2 = (xz)2 + (yz)2 . Facts (1), (2) and (3) together yield the “if” part of the theorem. (4) If p = 4m + 3 is a prime that divides a representable number n = x2 + y 2 , then p divides both x and y, and thus p2 divides n. In fact, if we had x ≡ 0 (mod p), then we could find x such that xx ≡ 1 (mod p), multiply the equation x2 + y 2 ≡ 0 by x2 , and thus obtain 1 + y 2 x2 = 1 + (xy)2 ≡ 0 (mod p), which is impossible for p = 4m + 3 by Lemma 1. (5) If n is representable, and p = 4m + 3 divides n, then p2 divides n, and n/p2 is representable. This follows from (4), and completes the proof. 

T On a finite set of odd cardinality, every involution has at least one fixed point.

22

Representing numbers as sums of two squares Two remarks close our discussion: • If a and b are two natural numbers that are relatively prime, then there are infinitely many primes of the form am + b (m ∈ N) — this is a famous (and difficult) theorem of Dirichlet. More precisely, one can show that the number of primes p ≤ x of the form p = am + b is described very x 1 accurately for large x by the function ϕ(a) log x , where ϕ(a) denotes the number of b with 1 ≤ b < a that are relatively prime to a. (This is a substantial refinement of the prime number theorem, which we had discussed on page 10.) • This means that the primes for fixed a and varying b appear essentially at the same rate. Nevertheless, for example for a = 4 one can observe a rather subtle, but still noticeable and persistent tendency towards “more” primes of the form 4m+3: If you look for a large random x, then chances are that there are more primes p ≤ x of the form p = 4m + 3 than of the form p = 4m + 1. This effect is known as “Chebyshev’s bias”; see Riesel [4] and Rubinstein and Sarnak [5].

References [1] F. W. C LARKE , W. N. E VERITT, L. L. L ITTLEJOHN & S. J. R. VORSTER : H. J. S. Smith and the Fermat Two Squares Theorem, Amer. Math. Monthly 106 (1999), 652-665. [2] D. R. H EATH -B ROWN : Fermat’s two squares theorem, Invariant (1984), 2-5. [3] I. N IVEN & H. S. Z UCKERMAN : An Introduction to the Theory of Numbers, Fifth edition, Wiley, New York 1972. [4] H. R IESEL : Prime Numbers and Computer Methods for Factorization, Second edition, Progress in Mathematics 126, Birkhäuser, Boston MA 1994. [5] M. RUBINSTEIN & P. S ARNAK : Chebyshev’s bias, Experimental Mathematics 3 (1994), 173-197. [6] A. T HUE : Et par antydninger til en taltheoretisk metode, Kra. Vidensk. Selsk. Forh. 7 (1902), 57-75. [7] S. WAGON : Editor’s corner: The Euclidean algorithm strikes again, Amer. Math. Monthly 97 (1990), 125-129. [8] D. Z AGIER : A one-sentence proof that every prime p ≡ 1 (mod 4) is a sum of two squares, Amer. Math. Monthly 97 (1990), 144.

The law of quadratic reciprocity

Chapter 5

Which famous mathematical theorem has been proved most often? Pythagoras would certainly be a good candidate or the fundamental theorem of algebra, but the champion is without doubt the law of quadratic reciprocity in number theory. In an admirable monograph Franz Lemmermeyer lists as of the year 2000 no fewer than 196 proofs. Many of them are of course only slight variations of others, but the array of different ideas is still impressive, as is the list of contributors. Carl Friedrich Gauss gave the first complete proof in 1801 and followed up with seven more. A little later Ferdinand Gotthold Eisenstein added five more — and the ongoing list of provers reads like a Who is Who of mathematics. With so many proofs present the question which of them belongs in the Book can have no easy answer. Is it the shortest, the most unexpected, or should one look for the proof that had the greatest potential for generalizations to other and deeper reciprocity laws? We have chosen two proofs (based on Gauss’ third and sixth proofs), of which the first may be the simplest and most pleasing, while the other is the starting point for fundamental results in more general structures. As in the previous chapter we work “modulo p”, where p is an odd prime; Zp is the field of residues upon division by p, and we usually (but not always) take these residues as 0, 1, . . . , p − 1. Consider some a ≡ 0 (mod p), that is, p  a. We call a a quadratic residue modulo p if a ≡ b2 (mod p) for some b, and a quadratic nonresidue otherwise. The quadratic residues are p−1 2 therefore 12 , 22 , . . . , ( p−1 2 ) , and so there are 2 quadratic residues and p−1 p−1 2 2 2 quadratic nonresidues. Indeed, if i ≡ j (mod p) with 1 ≤ i, j ≤ 2 , 2 2 then p | i − j = (i − j)(i + j). As 2 ≤ i + j ≤ p − 1 we have p | i − j, that is, i ≡ j (mod p). At this point it is convenient to introduce the so-called Legendre symbol. Let a ≡ 0 (mod p), then 

a 1 if a is a quadratic residue, := −1 if a is a quadratic nonresidue. p The story begins with Fermat’s “little theorem”: For a ≡ 0 (mod p), ap−1 ≡ 1 (mod p).

(1)

In fact, since Z∗p = Zp \ {0} is a group with multiplication, the set {1a, 2a, 3a, . . . , (p − 1)a} runs again through all nonzero residues, (1a)(2a) · · · ((p − 1)a) ≡ 1 · 2 · · · (p − 1) (mod p), and hence by dividing by (p − 1)!, we get ap−1 ≡ 1 (mod p). M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_5, © Springer-Verlag Berlin Heidelberg 2013

Carl Friedrich Gauss

For p = 13, the quadratic residues are 12 ≡ 1, 22 ≡ 4, 32 ≡ 9, 42 ≡ 3, 52 ≡ 12, and 62 ≡ 10; the nonresidues are 2, 5, 6, 7, 8, 11.

24

The law of quadratic reciprocity In other words, the polynomial xp−1 − 1 ∈ Zp [x] has as roots all nonzero residues. Next we note that xp−1 − 1 = (x

p−1 2

− 1)(x

p−1 2

+ 1).

Suppose a ≡ b2 (mod p) is a quadratic residue. Then by Fermat’s little p−1 theorem a 2 ≡ bp−1 ≡ 1 (mod p). Hence the quadratic residues are p−1 precisely the roots of the first factor x 2 − 1, and the p−1 2 nonresidues p−1 must thus be the roots of the second factor x 2 + 1. Comparing this to the definition of the Legendre symbol, we obtain the following important tool. For example, for p = 17 and a = 3 we have 38 = (34 )2 = 812 ≡ (−4)2 ≡ −1 (mod 17), while for a = 2 we get 28 = (24 )2 ≡ (−1)2 ≡ 1 (mod 17). Hence 2 is a quadratic residue, while 3 is a nonresidue.

Euler’s criterion. For a ≡ 0 (mod p),

a p−1 ≡ a 2 (mod p). p This gives us at once the important product rule

a  b 

ab  = , p p p

(2)

since this obviously holds for the right-hand side of Euler’s criterion. The product rule is extremely helpful when one tries to compute Legendre symbols: Since any integer is a product of ±1 and primes we only have to q 2 compute ( −1 p ), ( p ), and ( p ) for odd primes q. −1 By Euler’s criterion ( −1 p ) = 1 if p ≡ 1 (mod 4), and ( p ) = −1 if p ≡ 3 (mod 4), something we have already seen in the previous chapter. The case ( p2 ) will follow from the Lemma of Gauss below: ( p2 ) = 1 if p ≡ ±1 (mod 8), while ( p2 ) = −1 if p ≡ ±3 (mod 8). Euler, Legendre, and Gauss did lots of calculations with quadratic residues and, in particular, studied the relations between q being a quadratic residue modulo p and p being a quadratic residue modulo q, when p and q are odd primes. Euler and Legendre thus discovered the following remarkable theorem, but they managed to prove it only in special cases. However, Gauß was successful: On April 8, 1796 he was proud to record in his diary the first full proof.

Law of quadratic reciprocity. Let p and q be different odd primes. Then p−1 q−1 q p ( )( ) = (−1) 2 2 . p q

3 Example: ( 17 ) = ( 17 ) = ( 23 ) = −1, 3 so 3 is a nonresidue mod 17.

q−1 If p ≡ 1 (mod 4) or q ≡ 1 (mod 4), then p−1 2 (resp. 2 ) is even, and p−1 q−1 q p therefore (−1) 2 2 = 1; thus ( p ) = ( q ). When p ≡ q ≡ 3 (mod 4), we have ( pq ) = −( pq ). Thus for odd primes we get ( pq ) = ( qp ) unless both p and q are congruent to 3 (mod 4).

The law of quadratic reciprocity

25

First proof. The key to our first proof (which is Gauss’ third) is a counting formula that soon came to be called the Lemma of Gauss: Lemma of Gauss. Suppose a ≡ 0 (mod p). Take the numbers 1a, 2a, . . . , p−1 2 a and reduce them modulo p to the residue system p−1 smallest in absolute value, ia ≡ ri (mod p) with − p−1 2 ≤ ri ≤ 2 for all i. Then a ( ) = (−1)s , where s = #{i : ri < 0}. p  Proof. Suppose u1 , . . . , us are the residues smaller than 0, and that v1 , . . . , v p−1 −s are those greater than 0. Then the numbers −u1 , . . . , −us 2

are between 1 and p−1 2 , and are all different from the vj s (see the margin); hence {−u1 , . . . , −us , v1 , . . . , v p−1 −s } = {1, 2, . . . , p−1 2 }. Therefore 2

  (−ui ) vj = i

p−1 2 !,

j

which implies (−1)s

 i

ui



vj ≡

p−1 2 !

(mod p).

j

Now remember how we obtained the numbers ui and vj ; they are the residues of 1a, · · · , p−1 2 a. Hence   p−1 p−1 s 2 ui vj ≡ (−1)s p−1 (mod p). 2 ! ≡ (−1) 2 !a i

Cancelling

p−1 2 !

j

together with Euler’s criterion gives p−1 a ( ) ≡ a 2 ≡ (−1)s (mod p), p



and therefore ( ap ) = (−1)s , since p is odd.

With this we can easily compute ( p2 ): Since 1 · 2, 2 · 2, . . . , p−1 2 · 2 are all between 1 and p − 1, we have s = #{i :

p−1 2

< 2i ≤ p − 1} =

p−1 2

− #{i : 2i ≤

p−1 2 }

=  p−1 4 .

Check that s is even precisely for p = 8k ± 1. The Lemma of Gauss is the basis for many of the published proofs of the quadratic reciprocity law. The most elegant may be the one suggested by Ferdinand Gotthold Eisenstein, who had learned number theory from Gauss’ famous Disquisitiones Arithmeticae and made important contributions to “higher reciprocity theorems” before his premature death at age 29. His proof is just counting lattice points!

If −ui = vj , then ui + vj ≡ 0 (mod p). Now ui ≡ ka, vj ≡ a (mod p) implies p | (k + )a. As p and a are relatively prime, p must divide k +  which is impossible, since k +  ≤ p − 1.

26

The law of quadratic reciprocity Let p and q be odd primes, and consider ( pq ). Suppose iq is a multiple of q that reduces to a negative residue ri < 0 in the Lemma of Gauss. This means that there is a unique integer j such that − p2 < iq − jp < 0. Note that 0 < j < q2 since 0 < i < p2 . In other words, ( pq ) = (−1)s , where s is the number of lattice points (x, y), that is, pairs of integers x, y satisfying 0 < py − qx
0 for −1 < x < 1 and f (−1) = f (1) = 0. Then 2 2 T ≤ A ≤ R, 3 3 and equality holds in both cases only for n = 2. Erd˝os and Gallai established this result with an intricate induction proof. In the review of their paper, which appeared on the first page of the first issue of the Mathematical Reviews in 1940, George Pólya explained how the first inequality can also be proved by the inequality of the arithmetic and geometric mean — a beautiful example of a conscientious review and a Book Proof at the same time.

In praise of inequalities

123

 Proof of 23 T ≤ A. Since f (x) has only real roots, and none of them in the open interval (−1, 1), it can be written — apart from a constant positive factor which cancels out in the end — in the form   f (x) = (1 − x2 ) (αi − x) (βj + x) (3) i

j

with αi ≥ 1, βj ≥ 1. Hence 1 A =

(1 − x2 )

  (αi − x) (βj + x)dx. i

−1

j

By making the substitution x −→ −x, we find that also 1 A =

(1 − x2 )

  (αi + x) (βj − x)dx, j

i

−1

and hence by the inequality of the arithmetic and the geometric mean (note that all factors are ≥ 0) 1    1 (1 − x2 ) (αi − x) (βj + x) + A = 2 i j −1    (1 − x2 ) (αi + x) (βj − x) dx i

1 ≥

(1 − x2 )

 1/2  (α2i − x2 ) (βj2 − x2 ) dx i

−1

1 ≥

j

(1 − x2 )

j

 1/2  (α2i − 1) (βj2 − 1) dx i

−1

j

1/2  4  2 = (αi − 1) (βj2 − 1) . 3 i j Let us compute f (1) and f (−1). (We may assume f (−1), f (1) = 0, since otherwise T = 0 and the inequality 23 T ≤ A becomes trivial.) By (3) we see   f (1) = −2 (αi − 1) (βj + 1) , i

and similarly f (−1) = 2

j

  (αi + 1) (βj − 1) . i

j

Hence we conclude A ≥

2 (−f (1)f (−1))1/2 . 3

124

In praise of inequalities Applying now the inequality of the harmonic and the geometric mean to −f (1) and f (1), we arrive by (2) at the conclusion A ≥

2 3

1 −f  (1)

2 +

1 f  (−1)

=

2 4 f (1)f (−1) = T, 3 f (1) − f (−1) 3

which is what we wanted to show. By analyzing the case of equality in all our inequalities the reader can easily supply the last statement of the theorem.  The reader is invited to search for an equally inspired proof of the second inequality in Theorem 2. Well, analysis is inequalities after all, but here is an example from graph theory where the use of inequalities comes in quite unexpected. In Chapter 36 we will discuss Turán’s theorem. In the simplest case it takes on the following form. Theorem 3. Suppose G is a graph on n vertices without triangles. Then G 2 has at most n4 edges, and equality holds only when n is even and G is the complete bipartite graph Kn/2,n/2 .

i

 First proof. This proof, using Cauchy’s inequality, is due to Mantel. Let V = {1, . . . , n} be the vertex set and E the edge set of G. By di we denote

the degree of i, hence i∈V di = 2|E| (see page 165 in the chapter on double counting). Suppose ij is an edge. Since G has no triangles, we find di + dj ≤ n since no vertex is a neighbor of both i and j. It follows that  (di + dj ) ≤ n|E|.

j

ij∈E

...

...

Note that di appears exactly di times in the sum, so we get 

n|E| ≥

ij∈E

(di + dj ) =



d2i ,

i∈V

and hence with Cauchy’s inequality applied to the vectors (d1 , . . . , dn ) and (1, . . . , 1),

 4|E|2 ( di )2 2 n|E| ≥ = , di ≥ n n i∈V

and the result follows. In the case of equality we find di = dj for all i, j, and further di = n2 (since di + dj = n). Since G is triangle-free, G = Kn/2,n/2 is immediately seen from this. 

In praise of inequalities

125

 Second proof. The following proof of Theorem 3, using the inequality of the arithmetic and the geometric mean, is a folklore Book Proof. Let α be the size of a largest independent set A, and set β = n − α. Since G is triangle-free, the neighbors of a vertex i form an independent set, and we infer di ≤ α for all i. The set B = V \A of size β meets every edge of G. Counting

the edges of G according to their endvertices in B, we obtain |E| ≤ i∈B di . The inequality of the arithmetic and geometric mean now yields |E| ≤

 i∈B

di ≤ αβ ≤

 α + β 2 2

=

n2 , 4

and again the case of equality is easily dealt with.



References [1] H. A LZER : A proof of the arithmetic mean-geometric mean inequality, Amer. Math. Monthly 103 (1996), 585. [2] P. S. B ULLEN , D. S. M ITRINOVICS & P. M. VASI C´ : Means and their Inequalities, Reidel, Dordrecht 1988. ˝ & T. G RÜNWALD : On polynomials with only real roots, Annals [3] P. E RD OS Math. 40 (1939), 537-548.

[4] G. H. H ARDY, J. E. L ITTLEWOOD & G. P ÓLYA : Inequalities, Cambridge University Press, Cambridge 1952. [5] W. M ANTEL : Problem 28, Wiskundige Opgaven 10 (1906), 60-61. [6] G. P ÓLYA : Review of [3], Mathematical Reviews 1 (1940), 1. [7] G. P ÓLYA & G. S ZEG O˝ : Problems and Theorems in Analysis, Vol. I, SpringerVerlag, Berlin Heidelberg New York 1972/78; Reprint 1998.

i

"

... #$ di

%

The fundamental theorem of algebra

Chapter 19

Every nonconstant polynomial with complex coefficients has at least one root in the field of complex numbers. Gauss called this theorem, for which he gave four different proofs, the “fundamental theorem of algebraic equations.” It is without doubt one of the milestones in the history of mathematics. As Reinhold Remmert writes in his pertinent survey: “It was the possibility of proving this theorem in the complex domain that, more than anything else, paved the way for a general recognition of complex numbers.” Some of the greatest names have contributed to the subject, from Gauss and Cauchy to Liouville and Laplace. An article of Netto and Le Vavasseur lists nearly a hundred proofs. The proof that we present is one of the most elegant and certainly the shortest. It follows an argument of d’Alembert and Argand and uses only some elementary properties of polynomials and complex numbers. We are indebted to France Dacar for a polished version of the proof. Essentially the same argument appears also in the papers of Redheffer [4] and Wolfenstein [6], and doubtlessly in some others. We need three facts that one learns in a first-year calculus course. (A) Polynomial functions are continuous. (B) Any complex number of absolute value 1 has an m-th root for any m ≥ 1. (C) Cauchy’s minimum principle: A continuous real-valued function f on a compact set S assumes a minimum in S.

n Now let p(z) = k=0 ck z k be a complex polynomial of degree n ≥ 1. As the first and decisive step we prove what is variously called d’Alembert’s lemma or Argand’s inequality. Lemma. If p(a) = 0, then every disc D around a contains an interior point b with |p(b)| < |p(a)|.  Proof. Suppose the disc D has radius R. Thus the points in the interior of D are of the form a + w with |w| < R. First we will show that by a simple algebraic manipulation we may write p(a + w) as p(a + w) = p(a) + cwm (1 + r(w)),

(1)

where c is a nonzero complex number, 1 ≤ m ≤ n, and r(w) is a polynomial of degree n − m with r(0) = 0. M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_19, © Springer-Verlag Berlin Heidelberg 2013

It has been commented upon that the “Fundamental theorem of algebra” is not really fundamental, that it is not necessarily a theorem since sometimes it serves as a definition, and that in its classical form it is not a result from algebra, but rather from analysis.

Jean Le Rond d’Alembert

128

The fundamental theorem of algebra Indeed, we have p(a + w)

=

n 

ck (a + w)k

k=0

= =

n  n   k      k k−i i k a w = ck ak−i wi ck i i i=0 i=0 k=0 k=i   n n n   k  ck ak−i wi = p(a) + p(a) + di wi . i i=1 i=1 n 

k=i

Now let m ≥ 1 be the minimal index i for which di is different from zero, set c = dm , and factor cwm out to obtain p(a + w) = p(a) + cwm (1 + r(w)). Next wewant to bound |cwm | and |r(w)| from above. If |w| is smaller than ρ1 := m |p(a)/c|, then |cwm | < |p(a)|. Further, since r(w) is continuous and r(0) = 0, we have |r(w)| < 1 for |w| < ρ2 . Hence for |w| smaller than ρ := min(ρ1 , ρ2 ) we have |cwm | < |p(a)|

w0 a

b

and

|r(w)| < 1.

(2)

We come to our second ingredient, m-th roots of unity. Let ζ be an m-th p(a)/c root of − |p(a)/c| , which is a complex number of absolute value 1. Let ε be a real number with 0 < ε < min(ρ, R). Setting w0 = εζ we claim that b = a + w0 is a desired point in the disc D with |p(b)| < |p(a)|. First of all, b is in D, since |w0 | = ε < R, and further by (1), |p(b)| = |p(a + w0 )| = |p(a) + cw0m (1 + r(w0 ))|.

(3)

Now we define a factor δ by c w0m = c εm ζ m = −

εm p(a) = −δ p(a), |p(a)/c|

where by (2) our δ satisfies 0 < δ = εm

|c| < 1. |p(a)|

With the triangle inequality we get for the right-hand term in (3)

 |p(a) + cw0m 1 + r(w0 ) | = |p(a) − δp(a)(1 + r(w0 ))| = |(1 − δ)p(a) − δp(a)r(w0 )| ≤ < and we are done.

(1 − δ)|p(a)| + δ|p(a)||r(w0 )| (1 − δ)|p(a)| + δ|p(a)| = |p(a)|, 

The fundamental theorem of algebra The rest is easy. Clearly, p(z)z −n approaches the leading coefficient cn of p(z) as |z| goes to infinity. Hence |p(z)| goes to infinity as well with |z| → ∞. Consequently, there exists R1 > 0 such that |p(z)| > |p(0)| for all points z on the circle {z : |z| = R1 }. Furthermore, our third fact (C) tells us that in the compact set D1 = {z : |z| ≤ R1 } the continuous realvalued function |p(z)| attains the minimum value at some point z0 . Because of |p(z)| > |p(0)| for z on the boundary of D1 , z0 must lie in the interior. But by d’Alembert’s lemma this minimum value |p(z0 )| must be 0 — and this is the whole proof.

129

z0

R1 0

References [1] J. D ’A LEMBERT: Recherches sur le calcul intégral, Histoire de l’Académie Royale des Sciences et Belles Lettres (1746), 182-224. [2] R. A RGAND : Réflexions sur la nouvelle théorie d’analyse, Annales de Mathématiques 5 (1814), 197-209. [3] E. N ETTO & R. L E VAVASSEUR : Les fonctions rationelles, Enc. Sciences Math. Pures Appl. I 2 (1907), 1-232. [4] R. M. R EDHEFFER : What! Another note just on the fundamental theorem of algebra? Amer. Math. Monthly 71 (1964), 180-185. [5] R. R EMMERT: The fundamental theorem of algebra, Chapter 4 in: “Numbers” (H. D. Ebbinghaus et al., eds.), Graduate Texts in Mathematics 123, Springer, New York 1991. [6] S. W OLFENSTEIN : Proof of the fundamental theorem of algebra, Amer. Math. Monthly 74 (1967), 853-854.

“What’s up this time?” “Well, I’m shlepping 100 proofs for the Fundamental Theorem of Algebra”

“Proofs from the Book: one for the Fundamental Theorem, one for Quadratic Reciprocity!”

One square and an odd number of triangles

Chapter 20

Suppose we want to dissect a square into n triangles of equal area. When n is even, this is easily accomplished. For example, you could divide the horizontal sides into n2 segments of equal length and draw a diagonal in each of the n2 rectangles:

...

But now assume n is odd. Already for n = 3 this causes problems, and after some experimentation you will probably come to think that it might not be possible. So let us pose the general problem: Is it possible to dissect a square into an odd number n of triangles of equal area? Now, this looks like a classical question of Euclidean geometry, and one could have guessed that surely the answer must have been known for a long time (if not to the Greeks). But when Fred Richman and John Thomas popularized the problem in the 1960s they found to their surprise that no one knew the answer or a reference where this would be discussed. Well, the answer is “no” not only for n = 3, but for any odd n. But how should one prove a result like this? By scaling we may, of course, restrict ourselves to the unit square with vertices (0, 0), (1, 0), (0, 1), (1, 1). Any argument must therefore somehow make use of the fact that the area of the triangles in a dissection is n1 , where n is odd. The following proof due to Paul Monsky, with initial work of John Thomas, is a stroke of genius and totally unexpected: It uses an algebraic tool, valuations, to construct a striking coloring of the plane, and combines this with some elegant and stunningly simple combinatorial reasonings. And what’s more: at present no other proof is known! Before we state the theorem let us prepare the ground by a quick study of valuations. Everybody is familiar with the absolute value function |x| on the rationals Q (or the reals R). It maps Q to the nonnegative reals such that for all x and y, M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_20, © Springer-Verlag Berlin Heidelberg 2013

There are dissections of squares into an odd number of triangles whose areas are nearly equal.

132

One square and an odd number of triangles (i) |x| = 0 if and only if x = 0, (ii) |xy| = |x||y|, and (iii) |x + y| ≤ |x| + |y|

(the triangle inequality).

The triangle inequality makes R into a metric space and gives rise to the familiar notions of convergence. It was a great discovery around 1900 that besides the absolute value there are other natural “value functions” on Q that satisfy the conditions (i) to (iii). Let p be a prime number. Any rational number r = 0 can be written uniquely in the form a r = pk , k ∈ Z, (1) b where a and b > 0 are relatively prime to p. Define the p-adic value |r|p := p−k , | 34 |2

Example: = 4, | 76 |2 = |2|2 = 12 , and | 34 + 67 |2 = | 45 | = | 14 · 28 2 3 = 4 = max{| 4 |2 , | 67 |2 }.

45 | 7 2

|0|p = 0.

(2)

Conditions (i) and (ii) are obviously satisfied, and for (iii) we obtain the even stronger inequality (iii ) |x + y|p ≤ max{|x|p , |y|p }

(the non-Archimedean property).

Indeed, let r = pk ab and s = p dc , where we may assume that k ≥ , that is, |r|p = p−k ≤ p− = |s|p . Then we get & a & c && a c && & & |r + s|p = &pk + p & = &p (pk− + )& b d p b d p & pk− ad + bc & & & = p− & & ≤ p− = max{|r|p , |s|p }, bd p since the denominator bd is relatively prime to p. We also see from this that (iv) |x + y|p = max{|x|p , |y|p } whenever |x|p = |y|p , but we will prove below that this property is quite generally implied by (iii ). Any function v : K → R≥0 on a field K that satisfies (i) v(x) = 0 if and only if x = 0, (ii) v(xy) = v(x)v(y), and (iii ) v(x + y) ≤ max{v(x), v(y)}

(non-Archimedean property)

for all x, y ∈ K is called a non-Archimedean real valuation of K. For every such valuation v we have v(1) = v(1)v(1), hence v(1) = 1; and 1 = v(1) = v((−1)(−1)) = [v(−1)]2 , so v(−1) = 1. Thus from (ii) we get v(−x) = v(x) for all x and v(x−1 ) = v(x)−1 for x = 0. Every field has the trivial valuation that maps every nonzero element onto 1, and if v is a real non-Archimedean valuation, then so is v t for any positive real number t. So for Q we have the p-adic valuations and their powers, and a famous theorem of Ostrowski states that any nontrivial real nonArchimedean valuation of Q is of this form.

One square and an odd number of triangles

133

As announced, let us verify that the important property (iv) v(x + y) = max{v(x), v(y)} if v(x) = v(y) holds for any non-Archimedean valuation. Indeed, suppose that we have v(x) < v(y). Then v(y) = v((x + y) − x)

≤

max{v(x + y), v(x)} = v(x + y)

≤

max{v(x), v(y)} = v(y)

where (iii ) yields the inequalities, the first equality is clear, and the other two follow from v(x) < v(y). Thus v(x + y) = v(y) = max{v(x), v(y)}. Monsky’s beautiful approach to the square dissection problem used an extension of the 2-adic valuation |x|2 to a valuation v of R, where “extension” means that we require v(x) = |x|2 whenever x is in Q. Such a nonArchimedean real extension exists, but this is not standard algebra fare. In the following, we present Monsky’s argument in a version due to Hendrik Lenstra that requires much less; it only needs a valuation v that takes values in an arbitrary “ordered group”, not necessarily in (R>0 , · , 1. The definition and the existence of such a valuation will be provided in the appendix to this chapter. Here we just note that any valuation with v( 12 ) > 1 satisfies v( n1 ) = 1 for odd integers n. Indeed, v( 12 ) > 1 means that v(2) < 1, and thus v(2k) < 1 by (iii ) and induction on k. From this we get v(2k + 1) = 1 from (iv), and 1 thus again v( 2k+1 ) = 1 from (ii). Monsky’s Theorem. It is not possible to dissect a square into an odd number of triangles of equal area.

 Proof. In the following we construct a specific three-coloring of the plane with amazing properties. One of them is that the area of any triangle whose vertices have three different colors — which in the following is called a rainbow triangle — has a v-value larger than 1, so the area cannot be n1 for odd n. And then we verify that any dissection of the unit square must contain such a rainbow triangle, and the proof will be complete. The coloring of the points (x, y) of the real plane will be constructed by looking at the entries of the triple (x, y, 1) that have the maximal value under the valuation v. This maximum may occur once or twice or even three times. The color (blue, or green, or red) will record the coordinate of (x, y, 1) in which the maximal v-value occurs first: ⎧ ⎪ ⎨ blue if v(x) ≥ v(y), v(x) ≥ v(1), (x, y) is colored green if v(x) < v(y), v(y) ≥ v(1), ⎪ ⎩ red if v(x) < v(1), v(y) < v(1).

The property (iv) together with v(−x) = v(x) also implies that v(a ± b1 ± b2 ± · · · ± b ) = v(a) if v(a) > v(bi ) for all i.

134

One square and an odd number of triangles

(0, 1)

(0,

(1, 1)

This assigns a unique color to each point in the plane. The figure in the margin shows the color for each point in the unit square whose coordinates k are fractions of the form 20 . The following statement is the first step to the proof. Lemma 1. For any blue point pb = (xb , yb ), green point pg = (xg , yg ), and red point pr = (xr , yr ), the v-value of the determinant ⎛ ⎞ xb yb 1 det ⎝ xg yg 1 ⎠ xr yr 1

1 ) 2

is at least 1. (0, 0)

( 12 , 0)

(1, 0)  Proof. The determinant is a sum of six terms. One of them is the product

of the entries of the main diagonal, xb yg 1. By construction of the coloring each of the diagonal entries compared to the other entries in the row has a maximal v-value, so comparing with the last entry in each row (which is 1) we get v(xb yg 1) = v(xb )v(yg )v(1) ≥ v(1)v(1)v(1) = 1. Any of the other five summands of the determinant is a product of three matrix entries, one from each row (with a sign that as we know is irrelevant for the v-value). It picks at least one matrix entry below the main diagonal, whose v-value is strictly smaller than that of the diagonal entry in the same row, and at least one matrix entry above the main diagonal, whose v-value is not larger than that of the diagonal entry in the same row. Thus all of the five other summands of the determinant have a v-value that is strictly smaller than the summand corresponding to the main diagonal. Thus by property (iv) of non-Archimedean valuations, we find that the v-value of the determinant is given by the summand corresponding to the main diagonal, ⎛ ⎞ xb yb 1

 v det ⎝ xg yg 1 ⎠ = v(xb yg 1) ≥ 1. xr yr 1 

Corollary. Any line of the plane receives at most two different colors. The area of a rainbow triangle cannot be 0, and it cannot be n1 for odd n.  Proof. The area of the triangle with vertices at a blue point pb , a green point pg , and a red point pr is the absolute value of 1 2 ((xb

− xr )(yg − yr ) − (xg − xr )(yb − yr )),

which up to the sign is half the determinant of Lemma 1. The three points cannot lie on a line since the determinant cannot be 0, as v(0) = 0. The area of the triangle cannot be n1 , since in this case we would get ± n2 for the determinant, thus v(± n2 ) = v( 12 )−1 v( n1 ) < 1 because of v( 12 ) > 1 and v( n1 ) = 1, contradicting Lemma 1.



One square and an odd number of triangles And why did we construct this coloring? Because we are now going to show that in any dissection of the unit square S = [0, 1]2 into triangles (equal-sized or not!) there must always be a rainbow triangle, which according to the corollary cannot have area n1 for odd n. Thus the following lemma will complete the proof of Monsky’s theorem. Lemma 2. Every dissection of the unit square S = [0, 1]2 into finitely many triangles contains an odd number of rainbow triangles, and thus at least one.  Proof. The following counting argument is truly inspired. The idea is due to Emanuel Sperner, and will reappear with “Sperner’s Lemma” in Chapter 25.

Consider the segments between neighboring vertices in a given dissection. A segment is called a red-blue segment if one endpoint is red and the other is blue. For the example in the figure, the red-blue segments are drawn in purple. We make two observations, repeatedly using the fact from the corollary that on any line there can be points of at most two colors. (A) The bottom line of the square contains an odd number of red-blue segments, since (0, 0) is red and (1, 0) is blue, and all vertices in between are red or blue. So on the walk from the red end to the blue end of the bottom line, there must be an odd number of changes between red and blue. The other boundary lines of the square contain no red-blue segments. (B) If a triangle T has at most two colors at its vertices, then it contains an even number of red-blue segments on its boundary. However, every rainbow triangle has an odd number of red-blue segments on its boundary. Indeed, there is an odd number of red-blue segments between a red vertex and a blue vertex of a triangle, but an even number (if any) between any vertices with a different color combination. Thus a rainbow triangle has an odd number of red-blue segments in its boundary, while any other triangle has an even number (two or zero) of vertex pairs with the color combination red and blue. Now let us count the boundary red-blue segments summed over all triangles in the dissection. Since every red-blue segment in the interior of the square is counted twice, and there is an odd number on the boundary of S, this count is odd. Hence we conclude from (B) that there must be an odd number of rainbow triangles. 

135

136

One square and an odd number of triangles

Appendix: Extending valuations It is not at all obvious that an extension of a non-Archimedean real valuation from one field to a larger one is always possible. But it can be done, not only from Q to R, but generally from any field K to a field L that contains K. (This is known as “Chevalley’s theorem”; see for example the book by Jacobson [1].) In the following, we establish much less — but enough for our application to odd dissections. Indeed, in our proof for Monsky’s theorem we have not used the addition for values of v : R → R≥0 ; we have used only the multiplication and the order on R≥0 . Hence for our argument it is sufficient if the nonzero values of v lie in a (multiplicatively written) ordered abelian group (G, · , 0 , · , 1. So here is what we will establish: Theorem. The field of real numbers R has a non-Archimedean valuation to an ordered abelian group v : R → {0} ∪ G such that v( 12 ) > 1.  Proof. We first relate any valuation on a field to a subring of the field. (All the subrings that we consider contain 1.) Suppose v : K → {0} ∪ G is a valuation; let R := {x ∈ K : v(x) ≤ 1},

U := {x ∈ K : v(x) = 1}.

It is immediate that R is a subring of K, called the valuation ring corresponding to v. Furthermore, v(xx−1 ) = v(1) = 1 implies that v(x) = 1

One square and an odd number of triangles

137

if and only if v(x−1 ) = 1. Thus U is the set of units (invertible elements) of R. In particular, U is a subgroup of K × , where we write K × := K \{0} for the multiplicative group of K. Finally, with R−1 := {x−1 : x = 0} we have K = R ∪ R−1 . Indeed, if x ∈ R then v(x) > 1 and therefore v(x−1 ) < 1, thus x−1 ∈ R. The property K = R ∪ R−1 already characterizes all possible valuation rings in a given field. Lemma. A proper subring R ⊆ K is a valuation ring with respect to some valuation v into some ordered group G if and only if K = R ∪ R−1 .  Proof. We have seen one direction. Suppose now K = R ∪ R−1 . How should we construct the group G? If v : K → {0} ∪ G is a valuation corresponding to R, then v(x) < v(y) holds if and only if v(xy −1 ) < 1, that is, if and only if xy −1 ∈ R \ U . Also, v(x) = v(y) if and only if xy −1 ∈ U , or xU = yU as cosets in the factor group K × /U . Hence the natural way to proceed goes as follows. Take the quotient group G := K × /U , and define an order relation on G by setting xU < yU :⇐⇒ xy −1 ∈ R \ U. It is a nice exercise to check that this indeed makes G into an ordered group. The map v : K → {0} ∪ G is then defined in the most natural way: v(0) = 0,

and

v(x) := xU for x = 0.

It is easy to verify conditions (i) to (iii ) for v, and that R is the valuation ring corresponding to v.  In order to establish the theorem, it thus suffices to find a valuation ring B ⊂ R such that 12 ∈ / B. Claim. Any inclusion-maximal subring B ⊂ R with the property 12 ∈ / B is a valuation ring. First we should perhaps note that a maximal subring B ⊂ R with the property 21 ∈ / B exists. This is not quite trivial — but it does follow with a routine application of Zorn’s lemma, which is reviewed in the box. Indeed, if we have an ascending chain of subrings Bi ⊂ R that don’t contain 12 , then this chain has an upper bound, given by the union of all the subrings Bi , which again is a subring and does not contain 12 .

Zorn’s Lemma The Lemma of Zorn is of fundamental importance in algebra and other parts of mathematics when one wants to construct maximal structures. It also plays a decisive role in the logical foundations of mathematics. Lemma. Suppose P≤ is a nonempty partially ordered set with the property that every ascending chain (ai )≤ has an upper bound b, such that ai ≤ b for all i. Then P≤ contains a maximal element M , meaning that there is no c ∈ P with M < c.

Z ⊂ R is such a subring with but it is not maximal.

1 2

∈ / Z,

138

One square and an odd number of triangles To prove the Claim, let us assume that B ⊂ R is a maximal subring not containing 12 . If B is not a valuation ring, then there is some element α ∈ R\(B ∪ B −1 ). We denote by B[α] the subring generated by B ∪ α, that is, the set of all real numbers that can be written as polynomials in α with coefficients in B. Let 2B ⊆ B be the subset of all elements of the form 2b, for b ∈ B. Now 2B is a subset of B, so we have 2B[α] ⊆ B[α] and 2B[α−1 ] ⊆ B[α−1 ]. If we had 2B[α] = B[α] or 2B[α−1 ] = B[α−1 ], then due to 1 ∈ B this would imply that 12 ∈ / B[α] resp. 12 ∈ / B[α−1 ], contradicting the maximality of B ⊂ R as a subring that does not contain 12 . Thus we get that 2B[α] = B[α] and 2B[α−1 ] = B[α−1 ]. This implies that 1 ∈ B can be written in the form 1

= 2u0 + 2u1 α + · · · + 2um αm

with ui ∈ B,

1

= 2v0 + 2v1 α−1 + · · · + 2vn α−n with vi ∈ B, (2)

(1)

and similarly as which after multiplication by αn and subtraction of 2v0 αn from both sides yields (1 − 2v0 )αn

=

2v1 αn−1 + · · · + 2vn−1 α + 2vn .

(3)

Let us assume that these representations are chosen such that m and n are as small as possible. We may also assume that m ≥ n, otherwise we exchange α with α−1 , and (1) with (2). Now multiply (1) by 1 − 2v0 and add 2v0 on both sides of the equation, to get 1 = 2(u0 (1 − 2v0 ) + v0 ) + 2u1 (1 − 2v0 )α + · · · + 2um (1 − 2v0 )αm . But if in this equation we substitute for the term (1−2v0 )αm the expression given by equation (3) multiplied by αm−n , then this results in an equation that expresses 1 ∈ B as a polynomial in 2B[α] of degree at most m − 1. This contradiction to the minimality of m establishes the Claim. 

References [1] N. JACOBSON : Lectures in Abstract Algebra, Part III: Theory of Fields and Galois Theory, Graduate Texts in Mathematics 32. Springer, New York 1975. [2] P. M ONSKY: On dividing a square into triangles, Amer. Math. Monthly 77 (1970), 161-164. [3] F. R ICHMAN & J. T HOMAS : Problem 5471, Amer. Math. Monthly 74 (1967), 329. [4] S. K. S TEIN & S. S ZABÓ : Algebra and Tiling: Homomorphisms in the Service of Geometry, Carus Math. Monographs 25, MAA, Washington DC 1994. [5] J. T HOMAS : A dissection problem, Math. Magazine 41 (1968), 187-190.

Chapter 21

A theorem of Pólya on polynomials

Among the many contributions of George Pólya to analysis, the following has always been Erd˝os’ favorite, both for the surprising result and for the beauty of its proof. Suppose that f (z) = z n + bn−1 z n−1 + . . . + b0 is a complex polynomial of degree n ≥ 1 with leading coefficient 1. Associate with f (z) the set C := {z ∈ C : |f (z)| ≤ 2}, that is, C is the set of points which are mapped under f into the circle of radius 2 around the origin in the complex plane. So for n = 1 the domain C is just a circular disk of diameter 4. By an astoundingly simple argument, Pólya revealed the following beautiful property of this set C: Take any line L in the complex plane and consider the orthogonal projection CL of the set C onto L. Then the total length of any such projection never exceeds 4.

George Pólya

What do we mean by the total length of the projection CL being at most 4? We will see that CL is a finite union of disjoint intervals I1 , . . . , It , and the condition means that (I1 )+. . .+(It ) ≤ 4, where (Ij ) is the usual length of an interval. By rotating the plane we see that it suffices to consider the case when L is the real axis of the complex plane. With these comments in mind, let us state Pólya’s result. Theorem 1. Let f (z) be a complex polynomial of degree at least 1 and leading coefficient 1. Set C = {z ∈ C : |f (z)| ≤ 2} and let R be the orthogonal projection of C onto the real axis. Then there are intervals I1 , . . . , It on the real line which together cover R and satisfy (I1 ) + . . . + (It ) ≤ 4. Clearly the bound of 4 in the theorem is attained for n = 1. To get more of a feeling for the problem let us look at the polynomial f (z) = z 2 − 2, which also attains the bound of 4. If z = x + iy is a complex number, then x is its orthogonal projection onto the real line. Hence R = {x ∈ R : x + iy ∈ C for some y}. M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_21, © Springer-Verlag Berlin Heidelberg 2013

L I1

C

I2 ..

. It

C

..

. C

140

A theorem of Pólya on polynomials y

The reader can easily prove that for f (z) = z 2 − 2 we have x + iy ∈ C if and only if (x2 + y 2 )2 ≤ 4(x2 − y 2 ).

C R x

f (z) = z 2 − 2

z = x + iy

y

It follows that x4 ≤ (x2 + y 2 )2 ≤ 4x2 , and thus x2 ≤ 4, that is, |x| ≤ 2. On the other hand, any z = x ∈ R with |x| ≤ 2 satisfies |z 2 − 2| ≤ 2, and we find that R is precisely the interval [−2, 2] of length 4. As a first step towards the proof write f (z) = (z−c1 ) · · · (z−cn ) with ck = ak + ibk , and consider the real polynomial p(x) = (x − a1 ) · · · (x − an ). Let z = x + iy ∈ C, then by the theorem of Pythagoras |x − ak |2 + |y − bk |2 = |z − ck |2

ck = ak + ibk bk

and hence |x − ak | ≤ |z − ck | for all k, that is, |p(x)| = |x − a1 | · · · |x − an | ≤ |z − c1 | · · · |z − cn | = |f (z)| ≤ 2.

x

ak

Thus we find that R is contained in the set P = {x ∈ R : |p(x)| ≤ 2}, and if we can show that this latter set is covered by intervals of total length at most 4, then we are done. Accordingly, our main Theorem 1 will be a consequence of the following result. Theorem 2. Let p(x) be a real polynomial of degree n ≥ 1 with leading coefficient 1, and all roots real. Then the set P = {x ∈ R : |p(x)| ≤ 2} can be covered by intervals of total length at most 4. As Pólya shows in his paper [2], Theorem 2 is, in turn, a consequence of the following famous result due to Chebyshev. To make this chapter self-contained, we have included a proof in the appendix (following the beautiful exposition by Pólya and Szeg˝o). Chebyshev’s Theorem. Let p(x) be a real polynomial of degree n ≥ 1 with leading coefficient 1. Then 1 max |p(x)| ≥ n−1 . −1≤x≤1 2 Let us first note the following immediate consequence. Corollary. Let p(x) be a real polynomial of degree n ≥ 1 with leading coefficient 1, and suppose that |p(x)| ≤ 2 for all x in the interval [a, b]. Then b − a ≤ 4. 2  Proof. Consider the substitution y = b−a (x − a) − 1. This maps the x-interval [a, b] onto the y-interval [−1, 1]. The corresponding polynomial

q(y) = p( b−a 2 (y + 1) + a) n has leading coefficient ( b−a 2 ) and satisfies

Pavnuty Chebyshev on a Soviet stamp from 1946

max |q(y)| = max |p(x)|.

−1≤y≤1

a≤x≤b

A theorem of Pólya on polynomials

141

By Chebyshev’s theorem we deduce b−a n n 1 2 ≥ max |p(x)| ≥ ( b−a 2 ) 2n−1 = 2( 4 ) , a≤x≤b

and thus b − a ≤ 4, as desired.



This corollary brings us already very close to the statement of Theorem 2. If the set P = {x : |p(x)| ≤ 2} is an interval, then the length of P is at most 4. The set P may, however, not be an interval, as in the example depicted here, where P consists of two intervals. What can we say about P? Since p(x) is a continuous function, we know at any rate that P is the union of disjoint closed intervals I1 , I2 , . . ., and that p(x) assumes the value 2 or −2 at each endpoint of an interval Ij . This implies that there are only finitely many intervals I1 , . . . , It , since p(x) can assume any value only finitely often. Pólya’s wonderful idea was to construct another polynomial p˜(x) of degree ' = {x : |˜ n, again with leading coefficient 1, such that P p(x)| ≤ 2} is an interval of length at least (I1 ) + . . . + (It ). The corollary then proves ' ≤ 4, and we are done. (I1 ) + . . . + (It ) ≤ (P)  Proof of Theorem 2. Consider p(x) = (x − a1 ) · · · (x − an ) with P = {x ∈ R : |p(x)| ≤ 2} = I1 ∪ . . . ∪ It , where we arrange the intervals Ij such that I1 is the leftmost and It the rightmost interval. First we claim that any interval Ij contains a root of p(x). We know that p(x) assumes the values 2 or −2 at the endpoints of Ij . If one value is 2 and the other −2, then there is certainly a root in Ij . So assume p(x) = 2 at both endpoints (the case −2 being analogous). Suppose b ∈ Ij is a point where p(x) assumes its minimum in Ij . Then p (b) = 0 and p (b) ≥ 0. If p (b) = 0, then b is a multiple root of p (x), and hence a root of p(x) by Fact 1 from the box on the next page. If, on the other hand, p (b) > 0, then we deduce p(b) ≤ 0 from Fact 2 from the same box. Hence either p(b) = 0, and we have our root, or p(b) < 0, and we obtain a root in the interval from b to either endpoint of Ij . Here is the final idea of the proof. Let I1 , . . . , It be the intervals as before, and suppose the rightmost interval It contains m roots of p(x), counted with their multiplicities. If m = n, then It is the only interval (by what we just proved), and we are finished. So assume m < n, and let d be the distance between It−1 and It as in the figure. Let b1 , . . . , bm be the roots of p(x) which lie in It and c1 , . . . cn−m the remaining roots. We now write p(x) = q(x)r(x) where q(x) = (x − b1 ) · · · (x − bm ) and r(x) = (x − c1 ) · · · (x − cn−m ), and set p1 (x) = q(x + d)r(x). The polynomial p1 (x) is again of degree n with leading coefficient 1. For x ∈ I1 ∪. . .∪It−1 we have |x + d − bi | < |x − bi | for all i, and hence |q(x + d)| < |q(x)|. It follows that |p1 (x)| ≤ |p(x)| ≤ 2

for x ∈ I1 ∪ . . . ∪ It−1 .

If, on the other hand, x ∈ It , then we find |r(x − d)| ≤ |r(x)| and thus |p1 (x − d)| = |q(x)||r(x − d)| ≤ |p(x)| ≤ 2,

√ 1− 3

√ 1 1+ 3 ≈3.20

For the polynomial p(x) = x2 (x − 3) √ √ we get P = [1− 3, 1]∪[1+ 3, ≈ 3.2]

I1

... I2 . . .

d $ %" # It−1

It

142

A theorem of Pólya on polynomials which means that It − d ⊆ P1 = {x : |p1 (x)| ≤ 2}. In summary, we see that P1 contains I1 ∪. . .∪It−1 ∪(It −d) and hence has total length at least as large as P. Notice now that with the passage from p(x) to p1 (x) the intervals It−1 and It − d merge into a single interval. We conclude that the intervals J1 , . . . , Js of p1 (x) making up P1 have total length at least (I1 ) + . . . + (It ), and that the rightmost interval Js contains more than m roots of p1 (x). Repeating this procedure at most t − 1 ' = {x : |˜ times, we finally arrive at a polynomial p˜(x) with P p(x)| ≤ 2} ' being an interval of length (P) ≥ (I1 ) + . . . + (It ), and the proof is complete. 

Two facts about polynomials with real roots Let p(x) be a non-constant polynomial with only real roots. Fact 1. If b is a multiple root of p (x), then b is also a root of p(x).  Proof. Let

b1 < . . . < br be the roots of p(x) with multiplicities s1 , . . . , sr , rj=1 sj = n. From p(x) = (x − bj )sj h(x) we infer that bj is a root of p (x) if sj ≥ 2, and the multiplicity of bj in p (x) is sj − 1. Furthermore, there is a root of p (x) between b1 and b2 , another root between b2 and b3 , . . . , and one between br−1 and br , r and all these roots must be single roots, since j=1 (sj − 1) + (r − 1) counts already up to the degree n − 1 of p (x). Consequently, the multiple roots of p (x) can only occur among the roots of p(x).  Fact 2. We have p (x)2 ≥ p(x)p (x) for all x ∈ R.  Proof. If x = ai is a root of p(x), then there is nothing to show. Assume then x is not a root. The product rule of differentiation yields p (x) =

n  p(x) , x − ak

n  1 p (x) = . p(x) x − ak

that is,

k=1

k=1

Differentiating this again we have  1 p (x)p(x) − p (x)2 = − < 0. 2 p(x) (x − ak )2 n

k=1



A theorem of Pólya on polynomials

143

Appendix: Chebyshev’s theorem Theorem. Let p(x) be a real polynomial of degree n ≥ 1 with leading coefficient 1. Then 1 max |p(x)| ≥ n−1 . −1≤x≤1 2 Before we start, let us look at some examples where we have equality. The margin depicts the graphs of polynomials of degrees 1, 2 and 3, where we have equality in each case. Indeed, we will see that for every degree there is precisely one polynomial with equality in Chebyshev’s theorem.  Proof. Consider a real polynomial p(x) = xn + an−1 xn−1 + . . . + a0 with leading coefficient 1. Since we are interested in the range −1 ≤ x ≤ 1, we set x = cos ϑ and denote by g(ϑ) := p(cos ϑ) the resulting polynomial in cos ϑ, g(ϑ) = (cos ϑ)n + an−1 (cos ϑ)n−1 + . . . + a0 .

(1)

The proof proceeds now in the following two steps which are both classical results and interesting in their own right. (A) We express g(ϑ) as a so-called cosine polynomial, that is, a polynomial of the form g(ϑ) = bn cos nϑ + bn−1 cos(n − 1)ϑ + . . . + b1 cos ϑ + b0 with bk ∈ R, and show that its leading coefficient is bn =

(2)

1 2n−1 .

(B) Given any cosine polynomial h(ϑ) of order n (meaning that λn is the highest nonvanishing coefficient) h(ϑ) = λn cos nϑ + λn−1 cos(n − 1)ϑ + . . . + λ0 ,

(3)

we show |λn | ≤ max |h(ϑ)|, which when applied to g(ϑ) will then prove the theorem. Proof of (A). To pass from (1) to the representation (2), we have to express all powers (cos ϑ)k as cosine polynomials. For example, the addition theorem for the cosine gives cos 2ϑ = cos2 ϑ − sin2 ϑ = 2 cos2 ϑ − 1, so that cos2 ϑ = 12 cos 2ϑ + 12 . To do this for an arbitrary power (cos ϑ)k we go into the complex numbers, via the relation eix = cos x + i sin x. The eix are the complex numbers of absolute value 1 (see the box on complex unit roots on page 33). In particular, this yields einϑ = cos nϑ + i sin nϑ.

(4)

einϑ = (eiϑ )n = (cos ϑ + i sin ϑ)n .

(5)

On the other hand,

1 1 2 1 4

1

The polynomials p1 (x) = x, p2 (x) = x2 − 12 and p3 (x) = x3 − 34 x achieve equality in Chebyshev’s theorem.

144

A theorem of Pólya on polynomials Equating the real parts in (4) and (5) we obtain by i4+2 = −1, i4 = 1 and sin2 θ = 1 − cos2 θ n (cos ϑ)n−4 (1 − cos2 ϑ)2 cos nϑ = 4 ≥0 (6)  n  (cos ϑ)n−4−2 (1 − cos2 ϑ)2+1 . − 4 + 2 ≥0

We conclude that cos nϑ is a polynomial in cos ϑ, cos nϑ = cn (cos ϑ)n + cn−1 (cos ϑ)n−1 + . . . + c0 . 

n

= 2n−1 holds for n > 0: k≥0 2k Every subset of {1, 2, . . . , n − 1} yields an even sized subset of {1, 2, . . . , n} if we add the element n “if needed.”

(7)

From (6) we obtain for the highest coefficient n  n  + = 2n−1 . cn = 4 4 + 2 ≥0

≥0

Now we turn our argument around. Assuming by induction that for k < n, (cos ϑ)k can be expressed as a cosine polynomial of order k, we infer from (7) that (cos ϑ)n can be written as a cosine polynomial of order n with 1 leading coefficient bn = 2n−1 . Proof of (B). Let h(ϑ) be a cosine polynomial of order n as in (3), and assume without loss of generality λn > 0. Now we set m(ϑ) := λn cos nϑ and find m( nk π) = (−1)k λn for k = 0, 1, . . . , n. Suppose, for a proof by contradiction, that max |h(ϑ)| < λn . Then m( nk π) − h( nk π) = (−1)k λn − h( nk π) is positive for even k and negative for odd k in the range 0 ≤ k ≤ n. We conclude that m(ϑ) − h(ϑ) has at least n roots in the interval [0, π]. But this cannot be since m(ϑ) − h(ϑ) is a cosine polynomial of order n − 1, and thus has at most n − 1 roots. The proof of (B) and thus of Chebyshev’s theorem is complete.  The reader can now easily complete the analysis, showing that gn (ϑ) := 1 2n−1 cos nϑ is the only cosine polynomial of order n with leading coeffi1 cient 1 that achieves the equality max |g(ϑ)| = 2n−1 . The polynomials Tn (x) = cos nϑ, x = cos ϑ, are called the Chebyshev 1 polynomials (of the first kind); thus 2n−1 Tn (x) is the unique monic polynomial of degree n where equality holds in Chebyshev’s theorem.

References [1] P. L. C EBYCEV: Œuvres, Vol. I, Acad. Imperiale des Sciences, St. Petersburg 1899, pp. 387-469. [2] G. P ÓLYA : Beitrag zur Verallgemeinerung des Verzerrungssatzes auf mehrfach zusammenhängenden Gebieten, Sitzungsber. Preuss. Akad. Wiss. Berlin (1928), 228-232; Collected Papers Vol. I, MIT Press 1974, 347-351. [3] G. P ÓLYA & G. S ZEG O˝ : Problems and Theorems in Analysis, Vol. II, SpringerVerlag, Berlin Heidelberg New York 1976; Reprint 1998.

On a lemma of Littlewood and Offord

Chapter 22

In their work on the distribution of roots of algebraic equations, Littlewood and Offord proved in 1943 the following result: Let a1 , a2 , . . . , an be complex numbers

nwith |ai | ≥ 1 for all i, and consider the 2n linear combinations i=1 εi ai with εi ∈ {1, −1}.

n Then the number of sums i=1 εi ai which lie in the interior of any circle of radius 1 is not greater than 2n c √ log n n

for some constant c > 0.

A few years later Paul Erd˝os improved this bound by removing the log n term, but what is more interesting, he showed that this is, in fact, a simple consequence of the theorem of Sperner (see page 179). To get a feeling for his argument, let us look at the case when all ai are real. We may assume that all ai are positive (by changing ai to −ai

and εi to −εi whenever ai < 0). Now suppose that a set of combinations εi ai lies in the interior of

an interval of length 2. Let N = {1, 2, . . . , n} be the index set. For every εi ai we set I := {i ∈ N : εi = 1}. Now if I  I for two such sets, then we conclude that    ε i ai = 2 ai ≥ 2, ε i ai −

John E. Littlewood

i∈I  \I

which is a contradiction. Hence the sets I form an antichain,

n and  we conclude from the theorem of Sperner that there are at most n/2 such combinations. By Stirling’s formula (see page 11) we have   2n n for some c > 0. ≤ c√ n/2

n

n 

n For n even and all ai = 1 we obtain n/2 combinations i=1 εi ai that sum to 0. Looking at the interval (−1, 1) we thus find that the binomial number gives the exact bound.

n  In the same paper Erd˝os conjectured that n/2 was the right bound for

complex numbers as well (he could only prove c 2n n−1/2 for some c) and indeed that the same bound is valid for vectors a1 , . . . , an with |ai | ≥ 1 in a real Hilbert space, when the circle of radius 1 is replaced by an open ball of radius 1. M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_22, © Springer-Verlag Berlin Heidelberg 2013

Sperner’s theorem. Any antichain of subsets of an n-set has size at most  n . n/2

146

On a lemma of Littlewood and Offord Erd˝os was right, but it took twenty years until Gyula Katona and Daniel Kleitman independently came up with a proof for the complex numbers (or, what is the same, for the plane R2 ). Their proofs used explicitly the 2-dimensionality of the plane, and it was not at all clear how they could be extended to cover finite dimensional real vector spaces. But then in 1970 Kleitman proved the full conjecture on Hilbert spaces with an argument of stunning simplicity. In fact, he proved even more. His argument is a prime example of what you can do when you find the right induction hypothesis. A word of comfort for all readers who are not familiar with the notion of a Hilbert space: We do not really need general Hilbert spaces. Since we only deal with finitely many vectors ai , it is enough to consider the real space Rd with the usual scalar product. Here is Kleitman’s result. Theorem. Let a1 , . . . , an be vectors in Rd , each of length at least 1, and let R1 , . . . , Rk be k open regions of Rd , where |x − y| < 2 for any x, y that lie in the same

n region Ri . Then the number of linear combinations i=1 εi ai , εi ∈ {1, −1}, that can lie in the union i Ri of the regions is at most the sum of the k largest binomial coefficients nj .

n  In particular, we get the bound n/2 for k = 1. Before turning to the proof note that the bound is exact for a1 = . . . = an = a = (1, 0, . . . , 0)T .

n 

n  sums equal to 0, n/2−1 sums equal to Indeed, for even n we obtain n/2

n  (−2)a, n/2+1 sums equal to 2a, and so on. Choosing balls of radius 1 around −2 k−1 2 a,

...

(−2)a,

0,

2a,

...

2 k−1 2 a,

we obtain           n n n n n + . . . + n−2 + n + n+2 + . . . +



n+k−1 n−k+1 2 2 2 2 2 sums lying in these k balls, and this is our promised expression, since the largest binomial coefficients are centered around the middle (see page 12). A similar reasoning works when n is odd.  Proof. We may assume, without loss of generality, that the regions Ri are disjoint, and will do so from now on. The key to the proof is the recursion of the binomial coefficients, which tells us how the largest binomial n−k+1 n+k−1 coefficients

n nofn and nn− 1 are related. Set r = 2 , s = 2 , then r , r+1 , . . . , s are the k largest binomial coefficients for n. The

  n−1 recursion ni = n−1 + i−1 implies i

On a lemma of Littlewood and Offord s    n i=r

i

=

 s   n−1 i=r

=

 s   n−1 i=r

=

i

i

147

+

 s   n−1 i=r

i−1

 s−1   n−1 + i i=r−1

(1)

    s s−1   n−1 n−1 + , i i i=r−1 i=r

and an easy calculation that the first sum adds the k + 1 largest

shows  binomial coefficients n−1 , and the second sum the largest k − 1. i Kleitman’s proof proceeds by induction on n, the case n = 1 being trivial. In the light of (1) we need only show for the induction step that the linear combinations of a1 , . . . , an that lie in k disjoint regions can be mapped bijectively onto combinations of a1 , . . . , an−1 that lie in k + 1 or k − 1 regions. Claim. At least one of the translated regions Rj − an is disjoint from all the translated regions R1 + an , . . . , Rk + an . To prove this, consider the hyperplane H = {x : an , x = c} orthogonal to an , which contains all translates Ri + an on the side that is given by an , x ≥ c, and which touches the closure of some region, say Rj + an . Such a hyperplane exists since the regions are bounded. Now |x − y| < 2 holds for any x ∈ Rj and y in the closure of Rj , since Rj is open. We want to show that Rj − an lies on the other side of H. Suppose, on the contrary, that an , x − an  ≥ c for some x ∈ Rj , that is, an , x ≥ |an |2 + c. Let y + an be a point where H touches Rj + an , then y is in the closure of Rj , and an , y + an  = c, that is, an , −y = |an |2 − c. Hence an , x − y ≥ 2|an |2 , and we infer from the Cauchy–Schwarz inequality 2|an |2 ≤ an , x − y ≤ |an ||x − y|, and thus (with |an | ≥ 1) we get 2 ≤ 2|an | ≤ |x − y|, a contradiction.

The rest is easy. We classify the combinations

εi ai which come to lie in R1 ∪ . . .

∪ Rk as follows. Into Class 1 we put all ni=1 εi ai with εn = −1 n and all i=1 εi ai with εn = 1 lying in Rj , and into Class 2 we throw n in the remaining combinations i=1 εi ai with εn = 1, not in Rj . It

n−1 follows that the combinations i=1 εi ai corresponding to Class 1 lie in the k + 1 disjoint regions R1 + an , . . . , Rk + an and Rj − an , and the

n−1 combinations i=1 εi ai corresponding to Class 2 lie in the k − 1 disjoint regions R1 −a n, . . . , R

k −a  n without Rj −an . By induction, Class 1 con

s tains at most i=r−1 n−1 combinations, while Class 2 contains at most i

s−1 n−1 combinations — and by (1) this is the whole proof, straight i=r i from The Book. 

R2 + an

Rk + an

R1 + an y + an

Rj + an

Rj − an

an H

148

On a lemma of Littlewood and Offord

References ˝ : On a lemma of Littlewood and Offord, Bulletin Amer. Math. Soc. [1] P. E RD OS 51 (1945), 898-902.

[2] G. K ATONA : On a conjecture of Erd˝os and a stronger form of Sperner’s theorem, Studia Sci. Math. Hungar. 1 (1966), 59-63. [3] D. K LEITMAN : On a lemma of Littlewood and Offord on the distribution of certain sums, Math. Zeitschrift 90 (1965), 251-259. [4] D. K LEITMAN : On a lemma of Littlewood and Offord on the distributions of linear combinations of vectors, Advances Math. 5 (1970), 155-157. [5] J. E. L ITTLEWOOD & A. C. O FFORD : On the number of real roots of a random algebraic equation III, Mat. USSR Sb. 12 (1943), 277-285.

Cotangent and the Herglotz trick

Chapter 23

What is the most interesting formula involving elementary functions? In his beautiful article [2], whose exposition we closely follow, Jürgen Elstrodt nominates as a first candidate the partial fraction expansion of the cotangent function: ∞

π cot πx =

1  1  1 + + x n=1 x + n x − n

(x ∈ R\Z).

This elegant formula was proved by Euler in §178 of his Introductio in Analysin Infinitorum from 1748 and it certainly counts among his finest achievements. We can also write it even more elegantly as N  1 π cot πx = lim (1) N →∞ x+n n=−N

1 but one has to note that the evaluation of the sum n∈Z x+n is a bit dangerous, since the sum is only conditionally convergent, so its value depends on the “right” order of summation. We shall derive (1) by an argument of stunning simplicity which is attributed to Gustav Herglotz — the “Herglotz trick.” To get started, set f (x) := π cot πx,

g(x) :=

lim

N →∞

N  n=−N

Gustav Herglotz

1 , x+n

and let us try to derive enough common properties of these functions to see in the end that they must coincide . . .

f (x)

(A) The functions f and g are defined for all non-integral values and are continuous there. πx For the cotangent function f (x) = π cot πx = π cos sin πx , this is clear (see 1 1 the figure). For g(x), we first use the identity x+n + x−n = − n22x −x2 to rewrite Euler’s formula as ∞ 1  2x − . (2) π cot πx = x n=1 n2 − x2

π 1 4

1

Thus for (A) we have to prove that for every x ∈ / Z the series ∞  1 n=1

n2 − x2

converges uniformly in a neighborhood of x. M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_23, © Springer-Verlag Berlin Heidelberg 2013

The function f (x) = π cot πx

x

150

Cotangent and the Herglotz trick For this, we don’t get any problem with the first term, for n = 1, or with the terms with 2n − 1 ≤ x2 , since there is only a finite number of them. On the other hand, for n ≥ 2 and 2n − 1 > x2 , that is n2 − x2 > (n − 1)2 > 0, the summands are bounded by 0
r = |R|, and let f : N −→ R be a mapping. Then there exists some a ∈ R with |f −1 (a)| ≥ 2. We may even state a stronger inequality: There exists some a ∈ R with BnC |f −1 (a)| ≥ . (1) r In fact, we would have |f −1 (a)|
aj2 > · · · > ajn+1

(j1 < j2 < · · · < jn+1 )

of length n + 1, or both.

The reader may have fun in proving that for mn numbers the statement remains no longer true in general.

This time the application of the pigeon-hole principle is not immediate. Associate to each ai the number ti which is the length of a longest increasing subsequence starting at ai . If ti ≥ m + 1 for some i, then we have an increasing subsequence of length m + 1. Suppose then that ti ≤ m for all i. The function f : ai −→ ti mapping {a1 , . . . , amn+1 } to {1, . . . , m} tells us by (1) that there is some s ∈ {1, . . . , m} such that f (ai ) = s for mn m + 1 = n + 1 numbers ai . Let aj1 , aj2 , . . . , ajn+1 (j1 < · · · < jn+1 ) be these numbers. Now look at two consecutive numbers aji , aji+1 . If aji < aji+1 , then we would obtain an increasing subsequence of length s starting at aji+1 , and consequently an increasing subsequence of length s + 1 starting at aji , which cannot be since f (aji ) = s. We thus obtain a decreasing subsequence aj1 > aj2 > · · · > ajn+1 of length n + 1.  This simple-sounding result on monotone subsequences has a highly nonobvious consequence on the dimension of graphs. We don’t need here the notion of dimension for general graphs, but only for complete graphs Kn . It can be phrased in the following way. Let N = {1, . . . , n}, n ≥ 3, and consider m permutations π1 , . . . , πm of N . We say that the permutations πi represent Kn if to every three distinct numbers i, j, k there exists a permutation π in which k comes after both i and j. The dimension of Kn is then the smallest m for which a representation π1 , . . . , πm exists. As an example we have dim(K3 ) = 3 since any one of the three numbers must come last, as in π1 = (1, 2, 3), π2 = (2, 3, 1), π3 = (3, 1, 2). What

Pigeon-hole and double counting

163

about K4 ? Note first dim(Kn ) ≤ dim(Kn+1 ): just delete n + 1 in a representation of Kn+1 . So, dim(K4 ) ≥ 3, and, in fact, dim(K4 ) = 3, by taking π1 = (1, 2, 3, 4),

π2 = (2, 4, 3, 1),

π3 = (1, 4, 3, 2).

It is not quite so easy to prove dim(K5 ) = 4, but then, surprisingly, the dimension stays at 4 up to n = 12, while dim(K13 ) = 5. So dim(Kn ) seems to be a pretty wild function. Well, it is not! With n going to infinity, dim(Kn ) is, in fact, a very well-behaved function — and the key for finding a lower bound is the pigeon-hole principle. We claim dim(Kn ) ≥ log2 log2 n.

π1 : 1 2 3 5 6 7 8 9 10 11 12 4 π2 : 2 3 4 8 7 6 5 12 11 10 9 1 π3 : 3 4 1 11 12 9 10 6 5 8 7 2 π4 : 4 1 2 10 9 12 11 7 8 5 6 3 These four permutations represent K12

(2)

Since, as we have seen, dim(Kn ) is a monotone function in n, it suffices to p verify (2) for n = 22 + 1, that is, we have to show that dim(Kn ) ≥ p + 1

p

for n = 22 + 1.

Suppose, on the contrary, dim(Kn ) ≤ p, and let π1 , . . . , πp be representing p permutations of N = {1, 2, . . . , 22 + 1}. Now we use our result on monotone subsequences p times. In π1 there exists a monotone subsequence A1 p−1 of length 22 + 1 (it does not matter whether increasing or decreasing). Look at this set A1 in π2 . Using our result again, we find a monotone subp−2 + 1, and A2 is, of course, also sequence A2 of A1 in π2 of length 22 monotone in π1 . Continuing, we eventually find a subsequence Ap of size 0 22 + 1 = 3 which is monotone in all permutations πi . Let Ap = (a, b, c), then either a < b < c or a > b > c in all πi . But this cannot be, since there must be a permutation where b comes after a and c.  The right asymptotic growth was provided by Joel Spencer (upper bound) and by Füredi, Hajnal, Rödl and Trotter (lower bound): 1 dim(Kn ) = log2 log2 n + ( + o(1)) log2 log2 log2 n. 2 But this is not the whole story: Very recently, Morris and Ho¸sten found a method which, in principle, establishes the precise value of dim(Kn ). Using their result and a computer one can obtain the values given in the margin. This is truly astounding! Just consider how many permutations of size 1422564 there are. How does one decide whether 7 or 8 of them are required to represent K1422564 ?

3. Sums Paul Erd˝os attributes the following nice application of the pigeon-hole principle to Andrew Vázsonyi and Marta Sved: Claim. Suppose we are given n integers a1 , . . . , an , which need not be distinct. Then there is always

 a set of consecutive numbers ak+1 , ak+2 , . . . , a whose sum i=k+1 ai is a multiple of n.

dim(Kn ) ≤ 4 ⇐⇒ n ≤ 12 dim(Kn ) ≤ 5 ⇐⇒ n ≤ 81 dim(Kn ) ≤ 6 ⇐⇒ n ≤ 2646 dim(Kn ) ≤ 7 ⇐⇒ n ≤ 1422564

164

Pigeon-hole and double counting For the proof we set N = {0, 1, . . . , n} and R = {0, 1, . . . , n − 1}. Consider the map f : N → R, where f (m) is the remainder of a1 + . . . + am upon division by n. Since |N | = n + 1 > n = |R|, it follows that there are two sums a1 + . . . + ak , a1 + . . . + a (k < ) with the same remainder, where the first sum may be the empty sum denoted by 0. It follows that  

ai =

i=k+1

  i=1

ai −

k 

ai

i=1



has remainder 0 — end of proof.

Let us turn to the second principle: counting in two ways. By this we mean the following. Double counting. Suppose that we are given two finite sets R and C and a subset S ⊆ R × C. Whenever (p, q) ∈ S, then we say p and q are incident. If rp denotes the number of elements that are incident to p ∈ R, and cq denotes the number of elements that are incident to q ∈ C, then   rp = |S| = cq . (3) p∈R

q∈C

Again, there is nothing to prove. The first sum classifies the pairs in S according to the first entry, while the second sum classifies the same pairs according to the second entry. There is a useful way to picture the set S. Consider the matrix A = (apq ), the incidence matrix of S, where the rows and columns of A are indexed by the elements of R and C, respectively, with 9 1 if (p, q) ∈ S apq = 0 if (p, q) ∈ / S.

C 1 QQ R

1 2 3 4 5 6 7 8

2

3 4

5 6

7 8

1 1 1

1 1 1 1 1

1 1 1 1

1 1 1

With this set-up, rp is the sum of the p-th row of A and cq is the sum of the q-th column. Hence the first sum in (3) adds the entries of A (that is, counts the elements in S) by rows, and the second sum by columns. The following example should make this correspondence clear. Let R = C = {1, 2, . . . , 8}, and set S = {(i, j) : i divides j}. We then obtain the matrix in the margin, which only displays the 1’s.

1 1

4. Numbers again

1 1 1

Look at the table on the left. The number of 1’s in column j is precisely the number of divisors of j; let us denote this number by t(j). Let us ask how

Pigeon-hole and double counting

165

large this number t(j) is on the average when j ranges from 1 to n. Thus, we ask for the quantity 1 t(j). t¯(n) = n j=1

n

1 2 3 4 5 6 7 8

n

3 2

t¯(n) 1

5 3

2 2

7 3

16 7

5 2

The first few values of t¯(n)

How large is t¯(n) for arbitrary n? At first glance, this seems hopeless. For prime numbers p we have t(p) = 2, while for 2k we obtain a large number t(2k ) = k + 1. So, t(n) is a wildly jumping function, and we surmise that the same is true for t¯(n). Wrong guess, the opposite is true! Counting in two ways provides an unexpected and simple answer. Consider the matrix

n A (as above) for the integers 1 up to n. Counting by columns we get j=1 t(j). How many 1’s are in row i? Easy enough, the 1’s correspond to the multiples of i: 1i, 2i, . . ., and the last multiple not exceeding n is ni i. Hence we obtain n n n n  1 1n 1  n 1 ≤ = , t(j) = t¯(n) = n j=1 n i=1 i n i=1 i i i=1

where the error in each summand, when passing from ni to ni , is less than 1. Now the last sum is the n-th harmonic number Hn , so we obtain Hn − 1 < t¯(n) < Hn , and together with the estimates of Hn on page 11 this gives log n − 1 < Hn − 1 −

1 < t¯(n) < Hn < log n + 1. n

Thus we have proved the remarkable result that, while t(n) is totally erratic, the average t¯(n) behaves beautifully: It differs from log n by less than 1.

5. Graphs Let G be a finite simple graph with vertex set V and edge set E. We have defined in Chapter 12 the degree d(v) of a vertex v as the number of edges which have v as an end-vertex. In the example of the figure, the vertices 1, 2, . . . , 7 have degrees 3, 2, 4, 3, 3, 2, 3, respectively. Almost every book in graph theory starts with the following result (that we have already encountered in Chapters 12 and 18):  d(v) = 2|E|. (4) v∈V

For the proof consider S ⊆ V × E, where S is the set of pairs (v, e) such that v ∈ V is an end-vertex of e ∈ E. Counting S in two ways gives on the one hand v∈V d(v), since every vertex contributes d(v) to the count, and on the other hand 2|E|, since every edge has two ends.  As simple as the result (4) appears, it has many important consequences, some of which will be discussed as we go along. We want to single out in

6

1 4

2 3

5

7

166

Pigeon-hole and double counting this section the following beautiful application to an extremal problem on graphs. Here is the problem: Suppose G = (V, E) has n vertices and contains no cycle of length 4 (denoted by C4 ), that is, no subgraph . How many edges can G have at most? As an example, the graph in the margin on 5 vertices contains no 4-cycle and has 6 edges. The reader may easily show that on 5 vertices the maximal number of edges is 6, and that this graph is indeed the only graph on 5 vertices with 6 edges that has no 4-cycle. Let us tackle the general problem. Let G be a graph on n vertices without a 4-cycle. As above we denote by d(u) the degree of u. Now we count the following set S in two ways: S is the set of pairs (u, {v, w}) where u is adjacent to v and to w, with v = w. In other words, we count all u occurrences of v

w

d(u)

Summing over u, we find |S| = u∈V 2 . On the other hand, every pair {v, w} has at most one common neighbor (by the C4 -condition).  Hence |S| ≤ n2 , and we conclude    d(u) n ≤ 2 2 u∈V

or



d(u)2 ≤ n(n − 1) +

u∈V



d(u).

(5)

u∈V

Next (and this is quite typical for this sort of extremal problems) we apply the Cauchy–Schwarz inequality to the vectors (d(u1 ), . . . , d(un )) and (1, 1, . . . , 1), obtaining  2  d(u) ≤ n d(u)2 , u∈V

and hence by (5) 

u∈V

2  d(u) ≤ n2 (n − 1) + n d(u).

u∈V

u∈V

Invoking (4) we find 4 |E|2 ≤ n2 (n − 1) + 2n |E| or

n2 (n − 1) n |E| − ≤ 0. 2 4 Solving the corresponding quadratic equation we thus obtain the following result of Istvan Reiman. |E|2 −

Pigeon-hole and double counting Theorem. If the graph G on n vertices contains no 4-cycles, then n √  1 + 4n − 3 . |E| ≤ 4

167

(6)

For n = 5 this gives |E| ≤ 6, and the graph above shows that equality can hold. Counting in two ways has thus produced in an easy way an upper bound on the number of edges. But how good is the bound (6) in general? The following beautiful example [2] [3] [6] shows that it is almost sharp. As is often the case in such problems, finite geometry leads the way. In presenting the example we assume that the reader is familiar with the finite field Zp of integers modulo a prime p (see page 18). Consider the 3-dimensional vector space X over Zp . We construct from X the following graph Gp . The vertices of Gp are the one-dimensional subspaces [v] := spanZp {v}, 0 = v ∈ X, and we connect two such subspaces [v], [w] by an edge if

(0, 0, 1)

v, w = v1 w1 + v2 w2 + v3 w3 = 0. Note that it does not matter which vector = 0 we take from the subspace. In the language of geometry, the vertices are the points of the projective plane over Zp , and [w] is adjacent to [v] if w lies on the polar line of v. As an example, the graph G2 has no 4-cycle and contains 9 edges, which almost reaches the bound 10 given by (6). We want to show that this is true for any prime p. Let us first prove that Gp satisfies the C4 -condition. If [u] is a common neighbor of [v] and [w], then u is a solution of the linear equations v1 x + v2 y + v3 z = 0 w1 x + w2 y + w3 z = 0. Since v and w are linearly independent, we infer that the solution space has dimension 1, and hence that the common neighbor [u] is unique. Next, we ask how many vertices Gp has. It’s double counting again. The space X contains p3 − 1 vectors = 0. Since every one-dimensional sub3 −1 space contains p − 1 vectors = 0, we infer that X has pp−1 = p2 + p + 1 one-dimensional subspaces, that is, Gp has n = p2 + p + 1 vertices. Similarly, any two-dimensional subspace contains p2 −1 vectors = 0, and hence p2 −1 p−1 = p + 1 one-dimensional subspaces. It remains to determine the number of edges in Gp , or, what is the same by (4), the degrees. By the construction of Gp , the vertices adjacent to [u] are the solutions of the equation u1 x + u2 y + u3 z = 0.

(7)

The solution space of (7) is a two-dimensional subspace, and hence there are p + 1 vertices adjacent to [u]. But beware, it may happen that u itself is a solution of (7). In this case there are only p vertices adjacent to [u].

(1, 0, 1)

(0, 1, 1) (1, 1, 1)

(1, 0, 0)

(1, 1, 0)

(0, 1, 0)

The graph G2 : its vertices are all seven nonzero triples (x, y, z).

168

Pigeon-hole and double counting In summary, we obtain the following result: If u lies on the conic given by x2 + y 2 + z 2 = 0, then d([u]) = p, and, if not, then d([u]) = p + 1. So it remains to find the number of one-dimensional subspaces on the conic x2 + y 2 + z 2 = 0. Let us anticipate the result which we shall prove in a moment. Claim. There are precisely p2 solutions (x, y, z) of the equation x2 +y 2 +z 2 = 0, and hence (excepting the zero solution) precisely p2 −1 p−1 = p + 1 vertices in Gp of degree p. With this, we complete our analysis of Gp . There are p + 1 vertices of degree p, hence (p2 + p + 1) − (p + 1) = p2 vertices of degree p + 1. Using (4), we obtain |E|

= =

(p + 1)p p2 (p + 1) (p + 1)2 p + = 2 2 2  p2 + p (p + 1)p (1 + (2p + 1)) = (1 + 4p2 + 4p + 1). 4 4

Setting n = p2 + p + 1, the last equation reads |E| =

√ n−1 (1 + 4n − 3), 4

and we see that this almost agrees with (6).

⎛ 0 1 1 ⎜1 0 1 ⎜ ⎜1 1 0 ⎜ A=⎜ ⎜1 0 0 ⎜0 1 0 ⎜ ⎝0 0 1 0 0 0 The matrix for G2

1 0 0 1 0 0 1

0 1 0 0 1 0 1

0 0 1 0 0 1 1

⎞ 0 0⎟ ⎟ 0⎟ ⎟ 1⎟ ⎟ 1⎟ ⎟ 1⎠ 0

Now to the proof of the claim. The following argument is a beautiful application of linear algebra involving symmetric matrices and their eigenvalues. We will encounter the same method in Chapter 39, which is no coincidence: both proofs are from the same paper by Erd˝os, Rényi and Sós. We represent the one-dimensional subspaces of X as before by vectors v 1 , v 2 , . . . , v p2 +p+1 , any two of which are linearly independent. Similarly, we may represent the two-dimensional subspaces by the same set of vectors, where the subspace corresponding to u = (u1 , u2 , u3 ) is the set of solutions of the equation u1 x+u2 y+u3 z = 0 as in (7). (Of course, this is just the duality principle of linear algebra.) Hence, by (7), a one-dimensional subspace, represented by v i , is contained in the two-dimensional subspace, represented by v j , if and only if v i , v j  = 0. Consider now the matrix A = (aij ) of size (p2 +p+1)×(p2 +p+1), defined as follows: The rows and columns of A correspond to v 1 , . . . , v p2 +p+1 (we use the same numbering for rows and columns) with  1 if v i , v j  = 0, aij := 0 otherwise. A is thus a real symmetric matrix, and we have aii = 1 if v i , v i  = 0, that is, precisely when v i lies on the conic x2 + y 2 + z 2 = 0. Thus, all that remains to show is that trace A = p + 1.

Pigeon-hole and double counting

169

From linear algebra we know that the trace equals the sum of the eigenvalues. And here comes the trick: While A looks complicated, the matrix A2 is easy to analyze. We note two facts: • Any row of A contains precisely p + 1 1’s. This implies that p + 1 is an eigenvalue of A, since A1 = (p + 1)1, where 1 is the vector consisting of 1’s. • For any two distinct rows v i , v j there is exactly one column with a 1 in both rows (the column corresponding to the unique subspace spanned by v i , v j ). Using these facts we find ⎛ p+1 1 ⎜ ⎜ 1 p+1 A2 = ⎜ ⎜ . . ⎝ . 1 ···

··· ..

1 .. .

.

⎞ ⎟ ⎟ ⎟ = p I + J, ⎟ ⎠

p+1

where I is the identity matrix and J is the all-ones-matrix. Now, J has the eigenvalue p2 + p + 1 (of multiplicity 1) and 0 (of multiplicity p2 + p). Hence A2 has the eigenvalues p2 +2p+1 = (p+1)2 of multiplicity 1 and p of multiplicity p2 + p. Since A is real and symmetric, hence diagonalizable, we find that A has the eigenvalue p + 1 or −(p + 1) and p2 + p eigenvalues √ ± p. From Fact 1 above, the first eigenvalue must be p + 1. Suppose √ √ that p has multiplicity r, and − p multiplicity s, then √ √ trace A = (p + 1) + r p − s p. But now we are home: Since the trace is an integer, we must have r = s, so trace A = p + 1. 

6. Sperner’s Lemma In 1912, Luitzen Brouwer published his famous fixed point theorem: Every continuous function f : B n −→ B n of an n-dimensional ball to itself has a fixed point (a point x ∈ B n with f (x) = x). For dimension 1, that is for an interval, this follows easily from the intermediate value theorem, but for higher dimensions Brouwer’s proof needed some sophisticated machinery. It was therefore quite a surprise when in 1928 young Emanuel Sperner (he was 23 at the time) produced a simple combinatorial result from which both Brouwer’s fixed point theorem and the invariance of the dimension under continuous bijective maps could be deduced. And what’s more, Sperner’s ingenious lemma is matched by an equally beautiful proof — it is just double counting.

170

Pigeon-hole and double counting We discuss Sperner’s lemma, and Brouwer’s theorem as a consequence, for the first interesting case, that of dimension n = 2. The reader should have no difficulty to extend the proofs to higher dimensions (by induction on the dimension).

3 1 3

3

2

1

1

1

3 2

1

2

3

2

2

1

The triangles with three different colors are shaded

3 1 3

3

2

1

1

1

3 1

2

2

3

2

1

2

Sperner’s Lemma. Suppose that some “big” triangle with vertices V1 , V2 , V3 is triangulated (that is, decomposed into a finite number of “small” triangles that fit together edge-by-edge). Assume that the vertices in the triangulation get “colors” from the set {1, 2, 3} such that Vi receives the color i (for each i), and only the colors i and j are used for vertices along the edge from Vi to Vj (for i = j), while the interior vertices are colored arbitrarily with 1, 2 or 3. Then in the triangulation there must be a small “tricolored” triangle, which has all three different vertex colors.  Proof. We will prove a stronger statement: the number of tricolored triangles is not only nonzero, it is always odd. Consider the dual graph to the triangulation, but don’t take all its edges — only those which cross an edge that has endvertices with the (different) colors 1 and 2. Thus we get a “partial dual graph” which has degree 1 at all vertices that correspond to tricolored triangles, degree 2 for all triangles in which the two colors 1 and 2 appear, and degree 0 for triangles that do not have both colors 1 and 2. Thus only the tricolored triangles correspond to vertices of odd degree (of degree 1). However, the vertex of the dual graph which corresponds to the outside of the triangulation has odd degree: in fact, along the big edge from V1 to V2 , there is an odd number of changes between 1 and 2. Thus an odd number of edges of the partial dual graph crosses this big edge, while the other big edges cannot have both 1 and 2 occurring as colors. Now since the number of odd vertices in any finite graph is even (by equation (4)), we find that the number of small triangles with three different colors (corresponding to odd inside vertices of our dual graph) is odd.  With this lemma, it is easy to derive Brouwer’s theorem.  Proof of Brouwer’s fixed point theorem (for n = 2). Let Δ be the triangle in R3 with vertices e1 = (1, 0, 0), e2 = (0, 1, 0), and e3 = (0, 0, 1). It suffices to prove that every continuous map f : Δ −→ Δ has a fixed point, since Δ is homeomorphic to the two-dimensional ball B2 . We use δ(T ) to denote the maximal length of an edge in a triangulation T . One can easily construct an infinite sequence of triangulations T1 , T2 , . . . of Δ such that the sequence of maximal diameters δ(Tk ) converges to 0. Such a sequence can be obtained by explicit construction, or inductively, for example by taking Tk+1 to be the barycentric subdivision of Tk . For each of these triangulations, we define a 3-coloring of their vertices v by setting λ(v) := min{i : f (v)i < vi }, that is, λ(v) is the smallest index i such that the i-th coordinate of f (v) − v is negative. Assuming that f has no fixed point, this is well-defined. To see this, note that every v ∈ Δ lies

Pigeon-hole and double counting

in the plane x1 + x2 + x3 = 1, hence i vi = 1. So if f (v) = v, then at least one of the coordinates of f (v) − v must be negative (and at least one must be positive). Let us check that this coloring satisfies the assumptions of Sperner’s lemma. First, the vertex ei must receive color i, since the only possible negative component of f (ei ) − ei is the i-th component. Moreover, if v lies on the edge opposite to ei , then vi = 0, so the i-th component of f (v) − v cannot be negative, and hence v does not get the color i. Sperner’s lemma now tells us that in each triangulation Tk there is a tricolored triangle {v k:1 , v k:2 , v k:3 } with λ(v k:i ) = i. The sequence of points (v k:1 )k≥1 need not converge, but since the simplex Δ is compact some subsequence has a limit point. After replacing the sequence of triangulations Tk by the corresponding subsequence (which for simplicity we also denote by Tk ) we can assume that (v k:1 )k converges to a point v ∈ Δ. Now the distance of v k:2 and v k:3 from v k:1 is at most the mesh length δ(Tk ), which converges to 0. Thus the sequences (v k:2 ) and (v k:3 ) converge to the same point v. But where is f (v)? We know that the first coordinate f (v k:1 ) is smaller than that of v k:1 for all k. Now since f is continuous, we derive that the first coordinate of f (v) is smaller or equal to that of v. The same reasoning works for the second and third coordinates. Thus none of the coordinates of f (v) − v is positive — and we have already seen that this contradicts the assumption f (v) = v. 

References [1] L. E. J. B ROUWER : Über Abbildungen von Mannigfaltigkeiten, Math. Annalen 71 (1912), 97-115. [2] W. G. B ROWN : On graphs that do not contain a Thomsen graph, Canadian Math. Bull. 9 (1966), 281-285. ˝ , A. R ÉNYI & V. S ÓS : On a problem of graph theory, Studia Sci. [3] P. E RD OS Math. Hungar. 1 (1966), 215-235. ˝ & G. S ZEKERES : A combinatorial problem in geometry, Compositio [4] P. E RD OS Math. (1935), 463-470. ¸ & W. D. M ORRIS : The order dimension of the complete graph, [5] S. H O STEN Discrete Math. 201 (1999), 133-139.

[6] I. R EIMAN : Über ein Problem von K. Zarankiewicz, Acta Math. Acad. Sci. Hungar. 9 (1958), 269-273. [7] J. S PENCER : Minimal scrambling sets of simple orders, Acta Math. Acad. Sci. Hungar. 22 (1971), 349-353. [8] E. S PERNER : Neuer Beweis für die Invarianz der Dimensionszahl und des Gebietes, Abh. Math. Sem. Hamburg 6 (1928), 265-272. [9] W. T. T ROTTER : Combinatorics and Partially Ordered Sets: Dimension Theory, John Hopkins University Press, Baltimore and London 1992.

171

Tiling rectangles

Chapter 26

Some mathematical theorems exhibit a special feature: The statement of the theorem is elementary and easy, but to prove it can turn out to be a tantalizing task — unless you open some magic door and everything becomes clear and simple. One such example is the following result due to Nicolaas de Bruijn: Theorem. Whenever a rectangle is tiled by rectangles all of which have at least one side of integer length, then the tiled rectangle has at least one side of integer length. By a tiling we mean a covering of the big rectangle R with rectangles T1 , . . . , Tm that have pairwise disjoint interior, as in the picture to the right. Actually, de Bruijn proved the following result about packing copies of an a × b rectangle into a c × d rectangle: If a, b, c, d are integers, then each of a and b must divide one of c or d. This is implied by two applications of the more general theorem above to the given figure, scaled down first by a factor of a1 , and then scaled down by a factor of 1b . Each small rectangle has then one side equal to 1, and so ac or da must be an integer. Almost everybody’s first attempt is to try induction on the number of small rectangles. Induction can be made to work, but it has to be performed very carefully, and it is not the most elegant option one can come up with. Indeed, in a delightful paper Stan Wagon surveys no less than fourteen different proofs out of which we have selected three; none of them needs induction. The first proof, essentially due to de Bruijn himself, makes use of a very clever calculus trick. The second proof by Richard Rochberg and Sherman Stein is a discrete version of the first proof, which makes it simpler still. But the champion may be the third proof suggested by Mike Paterson. It is just counting in two ways and almost one-line. In the following we assume that the big rectangle R is placed parallel to the x, y-axes with (0, 0) as the lower left-hand corner. The small rectangles Ti have then sides parallel to the axes as well.  First Proof. Let T be any rectangle in the plane, where T extends from a to b along the x-axis and from c to d along the y-axis. Here is de Bruijn’s trick. Consider the double integral over T ,  d b e2πi(x+y) dx dy. (1) c

a

M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_26, © Springer-Verlag Berlin Heidelberg 2013

The big rectangle has side lengths 11 and 8.5.

174

Tiling rectangles Since 



d

b

e c

2πi(x+y)

 dx dy =

a

b

e

2πix



d

dx ·

a

e2πiy dy,

c

it follows that the integral (1) is 0 if and only if at least one of d or c e2πiy dy is equal to 0. We are going to show that 

b

e2πix dx = 0

⇐⇒

b − a is an integer.

b a

e2πix dx

(2)

a

 f (x, y) = R

  i

But  the tiling, each  then we will be done! Indeed, by the assumption on is equal to 0, and so by additivity of the integral, Ti R = 0 as well, whence R has an integer side. It remains to verify (2). From

f (x, y)

Ti

Additivity of the integral



b

e2πix dx

=

1 2πix &&b 1 2πib−2πia e (e ) & = 2πi 2πi a

=

e2πia 2πi(b−a) (e − 1) , 2πi

a

we conclude that 

b

e2πix dx = 0

⇐⇒

e2πi(b−a) = 1.

a

From e2πix = cos 2πx + i sin 2πx we see that the last equation is, in turn, equivalent to cos 2π(b − a) = 1 and sin 2π(b − a) = 0. Since cos x = 1 holds if and only if x is an integer multiple of 2π, we must have b − a ∈ Z, and this also implies sin 2π(b − a) = 0.  y

x

The amount of black in the corner rectangle is min(x, 12 ) · min(y, 12 ) + max(x − 12 , 0) · max(y − 12 , 0), and this is always greater than 12 xy.

 Second Proof. Color the plane in a checkerboard fashion with black/ white squares of size 21 × 12 , starting with a black square at (0, 0). By the assumption on the tiling every small rectangle Ti must receive an equal amount of black and white, and therefore the big rectangle R too contains the same amount of black and white. But this implies that R must have an integer side, since otherwise it can be split into four pieces, three of which have equal amounts of black and white, while the piece in the upper right-hand corner does not. Indeed, if x = a − a , y = b − b , so that 0 < x, y < 1, then the amount of black is always greater than that of white. This is illustrated in the figure in the margin. 

Tiling rectangles

175

 Third proof. Let C be the set of corners in the tiling for which both coordinates are integral (so, for example, (0, 0) ∈ C), and let T be the set of tiles. Form a bipartite graph G on the vertex set C ∪ T by joining each corner c ∈ C to all the tiles of which it is a corner. The hypothesis implies that each tile is joined to 0, 2, or 4 corners in C, since if one corner is in C, then so is also the other end of any integer side. G has therefore an even number of edges. Now look at C. Any c ∈ C which is not a corner of R is joined to an even number of tiles, but the vertex (0, 0) is joined to only one tile. So there must be another c ∈ C of odd degree, and c can only be one of the other vertices of R – end of proof. 

Here the bipartite graph G is drawn with vertices in C white, vertices in T black, and dashed edges.

All three proofs can quite easily be adapted to also yield an n-dimensional version of de Bruijn’s result: Whenever an n-dimensional box R is tiled by boxes all of which have at least one integer side, then R has an integer side. However, we want to keep our discussion in the plane (for this chapter), and look at a “companion result” to de Bruijn’s, due to Max Dehn (many years earlier), which sounds quite similar, but asks for different ideas. Theorem. A rectangle can be tiled with squares if and only if the ratio of its side lengths is a rational number. One half of the theorem is immediate. Suppose the rectangle R has side p α lengths α and β with α β ∈ Q, that is, β = q with p, q ∈ N. Setting

s

s := αp = βq , we can easily tile R with copies of the s × s square as shown β in the margin. For the proof of the converse Max Dehn used an elegant argument that he had already successfully employed in his solution of Hilbert’s third problem (see Chapter 9). In fact, the two papers appeared in successive years in the Mathematische Annalen.  Proof. Suppose R is tiled by squares of possibly different sizes. By scaling we may assume that R is an a × 1 rectangle. Let us assume a ∈ Q and derive a contradiction from this. The first step is to extend the sides of the squares to the full width resp. height of R as in the figure.

R is now decomposed into a number of small rectangles; let a1 , a2 , . . . , aM be their side lengths (in any order), and consider the set A := {1, a, a1 , . . . , aM } ⊂ R .

s

p squares

q squares

α

176

Tiling rectangles Next comes a linear algebra part. We define V (A) as the vector space of all linear combinations of the numbers in A with rational coefficients. Note that V (A) contains all side lengths of the squares in the original tiling, since any such side length is the sum of some ai ’s. As the number a is not rational, we may extend {1, a} to a basis B of V (A), B = {b1 = 1, b2 = a, b3 , . . . , bm } . Define the function f : B → R by f (1) := 1,

Linear extension: f (q1 b1 + · · · + qm bm ) := q1 f (b1 ) + · · · + qm f (bm ) for q1 , . . . , qm ∈ Q.

f (a) := −1,

and f (bi ) := 0 for i ≥ 3,

and extend it linearly to V (A). The following definition of “area” of rectangles finishes the proof in three quick steps: For c, d ∈ V (A) the area of the c × d rectangle is defined as area( (1) area (

c1 c2

d ) = area(

c c1

d ) = f (c)f (d). d ) + area(

c2

d ).

This follows immediately from the linearity of f . The analogous result holds, of course, for vertical strips.

(2) area(R) = area( ), where the sum runs through the squares in the tiling.

squares

Just note that by (1) area(R) equals the sum of the areas of all small rectangles in the extended tiling. Since any such rectangle is in exactly one square of the original tiling, we see (again by (1)) that this sum is also equal to the right-hand side of (2). (3) We have area(R) = f (a)f (1) = −1, whereas for a square of side length t, area( ) = f (t)2 ≥ 0, and so t  area( ) ≥ 0, squares

and this is our desired contradiction.



For those who want to go for further excursions into the world of tilings the beautiful survey paper [1] by Federico Ardila and Richard Stanley is highly recommended.

Tiling rectangles

177

References [1] F. A RDILA 32-43.

AND

R. P. S TANLEY: Tilings, Math. Intelligencer (4)32 (2010),

[2] N. G. DE B RUIJN : Filling boxes with bricks, Amer. Math. Monthly 76 (1969), 37-40. [3] M. D EHN : Über die Zerlegung von Rechtecken in Rechtecke, Mathematische Annalen 57 (1903), 314-332. [4] S. WAGON : Fourteen proofs of a result about tiling a rectangle, Amer. Math. Monthly 94 (1987), 601-617.

“The new hopscotch: Don’t hit the integers!”

Three famous theorems on finite sets

Chapter 27

In this chapter we are concerned with a basic theme of combinatorics: properties and sizes of special families F of subsets of a finite set N = {1, 2, . . . , n}. We start with two results which are classics in the field: the theorems of Sperner and of Erd˝os–Ko–Rado. These two results have in common that they were reproved many times and that each of them initiated a new field of combinatorial set theory. For both theorems, induction seems to be the natural method, but the arguments we are going to discuss are quite different and truly inspired. In 1928 Emanuel Sperner asked and answered the following question: Suppose we are given the set N = {1, 2, . . . , n}. Call a family F of subsets of N an antichain if no set of F contains another set of the family F . What is the size of a largest antichain? Clearly,

the family Fk of all k-sets satisfies the antichain property with |Fk | = nk . Looking at the maximum of the binomial (see page 12) we conclude that there is an antichain

coefficients   n of size n/2 = maxk nk . Sperner’s theorem now asserts that there are no larger ones.

n  Theorem 1. The size of a largest antichain of an n-set is n/2 .  Proof. Of the many proofs the following one, due to David Lubell, is probably the shortest and most elegant.  Let F be an arbitrary antichain. n Then we have to show |F | ≤ n/2 . The key to the proof is that we consider chains of subsets ∅ = C0 ⊂ C1 ⊂ C2 ⊂ · · · ⊂ Cn = N , where |Ci | = i for i = 0, . . . , n. How many chains are there? Clearly, we obtain a chain by adding one by one the elements of N , so there are just as many chains as there are permutations of N , namely n!. Next, for a set A ∈ F we ask how many of these chains contain A. Again this is easy. To get from ∅ to A we have to add the elements of A one by one, and then to pass from A to N we have to add the remaining elements. Thus if A contains k elements, then by considering all these pairs of chains linked together we see that there are precisely k!(n − k)! such chains. Note that no chain can pass through two different sets A and B of F , since F is an antichain. To

ncomplete the proof, let mk be the number of k-sets in F . Thus |F | = k=0 mk . Then it follows from our discussion that the number of chains passing through some member of F is n 

mk k! (n − k)!,

k=0

and this expression cannot exceed the number n! of all chains. Hence M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_27, © Springer-Verlag Berlin Heidelberg 2013

Emanuel Sperner

180

Three famous theorems on finite sets we conclude n  k=0

Check that the family of all n2 -sets for even n respectively the two families of all n−1 -sets and of all n+1 -sets, when 2 2 n is odd, are indeed the only antichains that achieve the maximum size!

mk

k!(n − k)! ≤ 1, n!

or

n  m

nk ≤ 1. k=0

k

Replacing the denominators by the largest binomial coefficient, we therefore obtain   n n   1 n

n  mk ≤ 1, that is, |F | = mk ≤ , n/2

n/2 k=0

and the proof is complete.

k=0



Our second result is of an entirely different nature. Again we consider the set N = {1, . . . , n}. Call a family F of subsets an intersecting family if any two sets in F have at least one element in common. It is almost immediate that the size of a largest intersecting family is 2n−1 . If A ∈ F , then the complement Ac = N \A has empty intersection with A and accordingly cannot be in F . Hence we conclude that an intersecting family contains at most half the number 2n of all subsets, that is, |F | ≤ 2n−1 . On the other hand, if we consider the family of all sets containing a fixed element, say the family F1 of all sets containing 1, then clearly |F1 | = 2n−1 , and the problem is settled. But now let us ask the following question: How large can an intersecting family F be if all sets in F have the same size, say k ? Let us call such families intersecting k-families. To avoid trivialities, we assume n ≥ 2k since otherwise any two k-sets intersect, and there is nothing to prove. Taking up the above idea, we certainly obtain such a family F1 by considering all k-sets containing a fixed element, say 1. Clearly, we obtain all sets in F1 by adding to 1 all (k − 1)-subsets of {2, 3, . . . , n}, hence |F1 | = n−1 k−1 . Can we do better? No — and this is the theorem of Erd˝os–Ko–Rado.

 Theorem 2. The largest size of an intersecting k-family in an n-set is n−1 k−1 when n ≥ 2k.

point

Paul Erd˝os, Chao Ko and Richard Rado found this result in 1938, but it was not published until 23 years later. Since then multitudes of proofs and variants have been given, but the following argument due to Gyula Katona is particularly elegant. edge

 Proof. The key to the proof is the following simple lemma, which at first sight seems to be totally unrelated to our problem. Consider a circle C divided by n points into n edges. Let an arc of length k consist of k + 1 consecutive points and the k edges between them. Lemma. Let n ≥ 2k, and suppose we are given t distinct arcs A1 , . . . , At of length k, such that any two arcs have an edge in common. Then t ≤ k.

A circle C for n = 6. The bold edges depict an arc of length 3.

To prove the lemma, note first that any point of C is the endpoint of at most one arc. Indeed, if Ai , Aj had a common endpoint v, then they would have

Three famous theorems on finite sets

181

to start in different direction (since they are distinct). But then they cannot have an edge in common as n ≥ 2k. Let us fix A1 . Since any Ai (i ≥ 2) has an edge in common with A1 , one of the endpoints of Ai is an inner point of A1 . Since these endpoints must be distinct as we have just seen, and since A1 contains k − 1 inner points, we conclude that there can be at most k − 1 further arcs, and thus at most k arcs altogether.  Now we proceed with the proof of the Erd˝os–Ko–Rado theorem. Let F be an intersecting k-family. Consider a circle C with n points and n edges as above. We take any cyclic permutation π = (a1 , a2 , . . . , an ) and write the numbers ai clockwise next to the edges of C. Let us count the number of sets A ∈ F which appear as k consecutive numbers on C. Since F is an intersecting family we see by our lemma that we get at most k such sets. Since this holds for any cyclic permutation, and since there are (n − 1)! cyclic permutations, we produce in this way at most k(n − 1)! sets of F which appear as consecutive elements of some cyclic permutation. How often do we count a fixed set A ∈ F ? Easy enough: A appears in π if the k elements of A appear consecutively in some order. Hence we have k! possibilities to write A consecutively, and (n − k)! ways to order the remaining elements. So we conclude that a fixed set A appears in precisely k!(n − k)! cyclic permutations, and hence that   k(n − 1)! (n − 1)! n−1 |F | ≤ = = .  k!(n − k)! (k − 1)!(n − 1 − (k − 1))! k−1 Again we may ask whether the families containing a fixed element are the only intersecting k-families of maximal size. This is certainly not true for n = 2k. For example, for n = 4 and k = 2 the family {1, 2}, {1, 3}, {2, 3}

 also has size 31 = 3. More generally, for n = 2k we get the largest

  intersecting k-families, of size 12 nk = n−1 k−1 , by arbitrarily including one out of every pair of sets formed by a k-set A and its complement N \A. But for n > 2k the special families containing a fixed element are indeed the only ones. The reader is invited to try his hand at the proof. Finally, we turn to the third result which is arguably the most important basic theorem in finite set theory, the “marriage theorem” of Philip Hall proved in 1935. It opened the door to what is today called matching theory, with a wide variety of applications, some of which we shall see as we go along. Consider a finite set X and a collection A1 , . . . , An of subsets of X (which need not be distinct). Let us call a sequence x1 , . . . , xn a system of distinct representatives of {A1 , . . . , An } if the xi are distinct elements of X, and if xi ∈ Ai for all i. Of course, such a system, abbreviated SDR, need not exist, for example when one of the sets Ai is empty. The content of the theorem of Hall is the precise condition under which an SDR exists.

1 3 2

4

An intersecting family for n = 4, k = 2

182

Three famous theorems on finite sets Before giving the result let us state the human interpretation which gave it the folklore name marriage theorem: Consider a set {1, . . . , n} of girls and a set X of boys. Whenever x ∈ Ai , then girl i and boy x are inclined to get married, thus Ai is just the set of possible matches of girl i. An SDR represents then a mass-wedding where every girl marries a boy she likes. Back to sets, here is the statement of the result. Theorem 3. Let A1 , . . . , An be a collection of subsets of a finite set X. Then there exists a system of distinct representatives if and only if the union of any m sets Ai contains at least m elements, for 1 ≤ m ≤ n. The condition is clearly necessary: If m sets Ai contain between them fewer than m elements, then these m sets can certainly not be represented by distinct elements. The surprising fact (resulting in the universal applicability) is that this obvious condition is also sufficient. Hall’s original proof was rather complicated, and subsequently many different proofs were given, of which the following one (due to Easterfield and rediscovered by Halmos and Vaughan) may be the most natural.

“A mass wedding”

 Proof. We use induction on n. For n = 1 there is nothing to prove. Let n > 1, and suppose {A1 , . . . , An } satisfies the condition of the theorem which we abbreviate by (H). Call a collection of  sets Ai with 1 ≤  < n a critical family if its union has cardinality . Now we distinguish two cases. Case 1: There is no critical family. Choose any element x ∈ An . Delete x from X and consider the collection A 1 , . . . , A n−1 with A i = Ai \{x}. Since there is no critical family, we find that the union of any m sets A i contains at least m elements. Hence by induction on n there exists an SDR x1 , . . . , xn−1 of {A 1 , . . . , A n−1 }, and together with xn = x, this gives an SDR for the original collection.

A

B

C

Case 2: There exists a critical family. D E

{B, C, D} is a critical family

After renumbering the sets we may assume that {A1 , . . . , A } is a critical ' with |X| ' = . Since  < n, we infer the exisfamily. Then i=1 Ai = X tence of an SDR for A1 , . . . , A by induction, that is, there is a numbering ' such that xi ∈ Ai for all i ≤ . x1 , . . . , x of X Consider now the remaining collection A+1 , . . . , An , and take any m of these sets. Since the union of A1 , . . . , A and these m sets contains at least  + m elements by condition (H), we infer that the m sets contain at least ' In other words, condition (H) is satisfied for the m elements outside X. family ' . . . , An \X. ' A+1 \X, ' CombinInduction now gives an SDR for A+1 , . . . , An that avoids X. ing it with x1 , . . . , x we obtain an SDR for all sets Ai . This completes the proof. 

Three famous theorems on finite sets As we mentioned, Hall’s theorem was the beginning of the now vast field of matching theory [6]. Of the many variants and ramifications let us state one particularly appealing result which the reader is invited to prove for himself: Suppose the sets A1 , . . . , An all have size k ≥ 1 and suppose further that no element is contained in more than k sets. Then there exist k SDR’s such that for any i the k representatives of Ai are distinct and thus together form the set Ai . A beautiful result which should open new horizons on marriage possibilities.

References [1] T. E. E ASTERFIELD : A combinatorial algorithm, J. London Math. Soc. 21 (1946), 219-226. ˝ , C. K O & R. R ADO : Intersection theorems for systems of finite sets, [2] P. E RD OS Quart. J. Math. (Oxford), Ser. (2) 12 (1961), 313-320.

[3] P. H ALL : On representatives of subsets, J. London Math. Soc. 10 (1935), 26-30. [4] P. R. H ALMOS & H. E. VAUGHAN : The marriage problem, Amer. J. Math. 72 (1950), 214-215. [5] G. K ATONA : A simple proof of the Erd˝os–Ko–Rado theorem, J. Combinatorial Theory, Ser. B 13 (1972), 183-184. [6] L. L OVÁSZ & M. D. P LUMMER : Matching Theory, Akadémiai Kiadó, Budapest 1986. [7] D. L UBELL : A short proof of Sperner’s theorem, J. Combinatorial Theory 1 (1966), 299. [8] E. S PERNER : Ein Satz über Untermengen einer endlichen Menge, Math. Zeitschrift 27 (1928), 544-548.

183

Shuffling cards

Chapter 28

How often does one have to shuffle a deck of cards until it is random? The analysis of random processes is a familiar duty in life (“How long does it take to get to the airport during rush-hour?”) as well as in mathematics. Of course, getting meaningful answers to such problems heavily depends on formulating meaningful questions. For the card shuffling problem, this means that we have • to specify the size of the deck (n = 52 cards, say), • to say how we shuffle (we’ll analyze top-in-at-random shuffles first, and then the more realistic and effective riffle shuffles), and finally • to explain what we mean by “is random” or “is close to random.” So our goal in this chapter is an analysis of the riffle shuffle, due to Edgar N. Gilbert and Claude Shannon (1955, unpublished) and Jim Reeds (1981, unpublished), following the statistician David Aldous and the former magician turned mathematician Persi Diaconis according to [1]. We will not reach the final precise result that 7 riffle shuffles are sufficient to get a deck of n = 52 cards very close to random, while 6 riffle shuffles do not suffice — but we will obtain an upper bound of 12, and we will see some extremely beautiful ideas on the way: the concepts of stopping rules and of “strong uniform time,” the lemma that strong uniform time bounds the variation distance, Reeds’ inversion lemma, and thus the interpretation of shuffling as “reversed sorting.” In the end, everything will be reduced to two very classical combinatorial problems, namely the coupon collector and the birthday paradox. So let’s start with these!

The birthday paradox Take n random people — the participants of a class or seminar, say. What is the probability that they all have different birthdays? With the usual simplifying assumptions (365 days a year, no seasonal effects, no twins present) the probability is p(n) =

n−1  i=1

1−

i  , 365

M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_28, © Springer-Verlag Berlin Heidelberg 2013

Persi Diaconis’ business card as a magician. In a later interview he said: “If you say that you are a professor at Stanford people treat you respectfully. If you say that you invent magic tricks, they don’t want to introduce you to their daughter.”

186

Shuffling cards which is smaller than 12 for n = 23 (this is the “birthday paradox”!), less than 9 percent for n = 42, and exactly 0 for n > 365 (the “pigeon-hole principle,” see Chapter 25). The formula is easy to see — if we take the persons in some fixed order: If the first i persons have distinct birthdays, then the probability that the (i + 1)-st person doesn’t spoil the series is i 1 − 365 , since there are 365 − i birthdays left. Similarly, if n balls are placed independently and randomly into K boxes, then the probability that no box gets more than one ball is p(n, K) =

n−1  i=1

1−

i  . K

The coupon collector



Children buy photos of pop stars (or soccer stars) for their albums, but they buy them in little non-transparent envelopes, so they don’t know which photo they will get. If there are n different photos, what is the expected number of pictures a kid has to buy until he or she gets every motif at least once? Equivalently, if you randomly take balls from a bowl that contains n distinguishable balls, and if you put your ball back each time, and then again mix well, how often do you have to draw on average until you have drawn each ball at least once? If you already have drawn k distinct balls, then the probability not to get a new one in the next drawing is nk . So the probability to need exactly s drawings for the next new ball is ( nk )s−1 (1 − nk ). And thus the expected number of drawings for the next new ball is

xs−1 (1 − x)s =

s≥1

=



xs−1 s −

s≥1

=



s≥1

xs (s + 1) −

s≥0

=

 s≥0



xs s 

  k s−1 

xs s

s≥1

s≥0

xs =

1 , 1−x

where at the end we sum a geometric series (see page 36).

n

1−

1 k s = n 1−

k n

,

as we get from the series in the margin. So the expected number of drawings until we have drawn each of the n different balls at least once is n−1  k=0

1 1−

k n

=

n n n n + + ··· + + = nHn ≈ n log n, n n−1 2 1

with the bounds on the size of harmonic numbers that we had obtained on page 11. So the answer to the coupon collector’s problem is that we have to expect that roughly n log n drawings are necessary. The estimate that we need in the following is for the probability that you need significantly more than n log n trials. If Vn denotes the number of drawings needed (this is the random variable whose expected value is E[Vn ] ≈ n log n), then for n ≥ 1 and c ≥ 0, the probability that we need more than m := n log n + cn drawings is ( ) Prob Vn > m ≤ e−c .

Shuffling cards

187

Indeed, if Ai denotes the event that the ball i is not drawn in the first m drawings, then ( ) Prob Vn > m =

i

=

 n A little calculus shows that 1 − n1 is an increasing function in n, which con n verges to 1/e. So 1 − n1 < 1e holds for all n ≥ 1.

   ( ) Prob Ai ≤ Prob Ai i

 1 m n 1− < ne−m/n ≤ e−c . n

Now let’s grab a deck of n cards. We number them 1 up to n in the order in which they come — so the card numbered “1” is at the top of the deck, while “n” is at the bottom. Let us denote from now on by Sn the set of all permutations of 1, . . . , n. Shuffling the deck amounts to the application of certain random permutations to the order of the cards. Ideally, this might mean that we apply an arbitrary permutation π ∈ Sn to 1 our starting order (1, 2, . . . , n), each of them with the same probability n! . Thus, after doing this just once, we would have our deck of cards in order π = (π(1), π(2), . . . , π(n)), and this would be a perfect random order. But that’s not what happens in real life. Rather, when shuffling only “certain” permutations occur, perhaps not all of them with the same probability, and this is repeated a “certain” number of times. After that, we expect or hope the deck to be at least “close to random.”

Top-in-at-random shuffles These are performed as follows: you take the top card from the deck, and insert it into the deck at one of the n distinct possible places, each of them with probability n1 . Thus one of the permutations i ↓ 

τi = 2, 3, . . . , i, 1, i+1, . . . , n is applied, 1 ≤ i ≤ n. After one such shuffle the deck doesn’t look random, and indeed we expect to need lots of such shuffles until we reach that goal. A typical run of top-in-at-random shuffles may look as follows (for n = 5):

“Top-in-at-random”

1

2

3

2

2

4

2 3

3 4

2 4

4 1

4 1

1 5

4 5

1 5

1 5

5 3

5 3

3 2

How should we measure “being close to random”? Probabilists have cooked up the “variation distance” as a rather unforgiving measure of randomness: We look at the probability distribution on the n! different orderings of our deck, or equivalently, on the n! different permutations σ ∈ Sn that yield the orderings.

...

188

Shuffling cards Two examples are our starting distribution E, which is given by E(id) = E(π) =

1, 0 otherwise,

and the uniform distribution U given by U(π) =

1 n!

for all π ∈ Sn .

The variation distance between two probability distributions Q1 and Q2 is now defined as  Q1 − Q2  := 12 |Q1 (π) − Q2 (π)|. π∈Sn

By

setting S := {π ∈ Sn : Q1 (π) > Q2 (π)} and using π Q2 (π) = 1 we can rewrite this as Q1 − Q2  =

π

Q1 (π) =

max |Q1 (S) − Q2 (S)|,

S⊆Sn

with Qi (S) := π∈S Qi (π). Clearly we have 0 ≤ Q1 − Q2  ≤ 1. In the following, “being close to random” will be interpreted as “having small variation distance from the uniform distribution.” Here the distance between the starting distribution and the uniform distribution is very close to 1: 1 E − U = 1 − n! . After one top-in-at-random shuffle, this will not be much better: Top − U = 1 − For card players, the question is not “exactly how close to uniform is the deck after a million riffle shuffles?”, but “is 7 shuffles enough?” (Aldous & Diaconis [1])

d(k)

1

k0

1 (n−1)! .

The probability distribution on Sn that we obtain by applying the top-in-atrandom shuffle k times will be denoted by Top∗k . So how does Top∗k −U behave if k gets larger, that is, if we repeat the shuffling? And similarly for other types of shuffling? General theory (in particular, Markov chains on finite groups; see e. g. Behrends [3]) implies that for large k the variation distance d(k) := Top∗k − U goes to zero exponentially fast, but it does not yield the “cut-off” phenomenon that one observes in practice: After a certain number k0 of shuffles “suddenly” d(k) goes to zero rather fast. Our margin displays a schematic sketch of the situation.

Strong uniform stopping rules The amazing idea of strong uniform stopping rules by Aldous and Diaconis captures the essential features. Imagine that the casino manager closely watches the shuffling process, analyzes the specific permutations that are applied to the deck in each step, and after a number of steps that depends on the permutations that he has seen calls “STOP!”. So he has a stopping rule that ends the shuffling process. It depends only on the (random) shuffles that have already been applied. The stopping rule is strong uniform if the following condition holds for all k ≥ 0: If the process is stopped after exactly k steps, then the resulting permutations of the deck have uniform distribution (exactly!).

Shuffling cards

189

Let T be the number of steps that are performed until the stopping rule tells the manager to cry “STOP!”; so this is a random variable. Similarly, the ordering of the deck after k shuffles is given by a random variable Xk (with values in Sn ). With this, the stopping rule is strong uniform if for all feasible values of k, ( ) 1 Prob Xk = π | T = k = n!

Conditional probabilities The conditional probability Prob[A | B] denotes the probability of the event A under the condition that B happens. This is just the probability that both events happen, divided by the probability that B is true, that is,

for all π ∈ Sn .

Three aspects make this interesting, useful, and remarkable: 1. Strong uniform stopping rules exist: For many examples they are quite simple.

Prob[A | B] =

2. Moreover, these can be analyzed: Trying to determine Prob[T > k] leads often to simple combinatorial problems.

Prob[A ∧ B] . Prob[B]

3. This yields effective upper bounds on the variation distances such as d(k) = Top∗k − U. For example, for the top-in-at-random shuffles a strong uniform stopping rule is “STOP after the original bottom card (labelled n) is first inserted back into the deck.” Indeed, if we trace the card n during these shuffles, 1

2

3

2

2

4

2 3

3 4

2 4

4 1

4 1

1 5

4 5

1 5

1 5

5 3

5 3

3 2

T1 we see that during the whole process the ordering of the cards below this card is completely uniform. So, after the card n rises to the top and then is inserted at random, the deck is uniformly distributed; we just don’t know when precisely this happens (but the manager does). Now let Ti be the random variable which counts the number of shuffles that are performed until for the first time i cards lie below card n. So we have to determine the distribution of T = T1 + (T2 − T1 ) + . . . + (Tn−1 − Tn−2 ) + (T − Tn−1 ). But each summand in this corresponds to a coupon collector’s problem: Ti − Ti−1 is the time until the top card is inserted at one of the i possible places below the card n. So it is also the time that the coupon collector takes from the (n − i)-th coupon to the (n − i + 1)-st coupon. Let Vi be the number of pictures bought until he has i different pictures. Then Vn = V1 + (V2 − V1 ) + . . . + (Vn−1 − Vn−2 ) + (Vn − Vn−1 ),

T2

...

190

Shuffling cards and we have seen that Prob[Ti − Ti−1 = j] = Prob[Vn−i+1 − Vn−i = j] for all i and j. Hence the coupon collector and the top-in-at-random shuffler perform equivalent sequences of independent random processes, just in the opposite order (for the coupon collector, it’s hard at the end). Thus we know that the strong uniform stopping rule for the top-in-at-random shuffles takes more than k = n log n + cn steps with low probability: ( ) Prob T > k ≤ e−c . And this in turn means that after k = n log n + cn top-in-at-random shuffles, our deck is “close to random,” with d(k) = Top∗k − U ≤ e−c , due to the following simple but crucial lemma. Lemma. Let Q : Sn −→ R be any probability distribution that defines a shuffling process Q∗k with a strong uniform stopping rule whose stopping time is T . Then for all k ≥ 0, ( ) Q∗k − U ≤ Prob T > k .  Proof. If X is a random variable with values in Sn , with probability distribution Q, then we write Q(S) for the probability that X takes a value in S ⊆ Sn . Thus Q(S) = Prob[X ∈ S], and in the case of the uniform distribution Q = U we get ( ) |S| U(S) = Prob X ∈ S = . n! For every subset S ⊆ Sn , we get the probability that after k steps our deck is ordered according to a permutation in S as Q∗k (S) = Prob[Xk ∈ S]  = Prob[Xk ∈ S ∧ T = j] + Prob[Xk ∈ S ∧ T > k] j≤k

=



U(S) Prob[T = j] + Prob[Xk ∈ S | T > k] · Prob[T > k]

j≤k

= =

U(S) (1 − Prob[T > k]) + Prob[Xk ∈ S | T > k] · Prob[T > k] 

U(S) + Prob[Xk ∈ S | T > k] − U(S) · Prob[T > k].

This yields

|Q∗k (S) − U(S)| ≤ Prob[T > k]

since Prob[Xk ∈ S | T > k] − U(S) is a difference of two probabilities, so it has absolute value at most 1.



This is the point where we have completed our analysis of the top-in-atrandom shuffle: We have proved the following upper bound for the number of shuffles needed to get “close to random.”

Shuffling cards

191

Theorem 1. Let c ≥ 0 and k := n log n + cn. Then after performing k top-in-at-random shuffles on a deck of n cards, the variation distance from the uniform distribution satisfies d(k) := Top∗k − U ≤ e−c . One can also verify that the variation distance d(k) stays large if we do significantly fewer than n log n top-in-at-random shuffles. The reason is that a smaller number of shuffles will not suffice to destroy the relative ordering on the lowest few cards in the deck. Of course, top-in-at-random shuffles are extremely ineffective — with the bounds of our theorem, we need more than n log n ≈ 205 top-in-at random shuffles until a deck of n = 52 cards is mixed up well. Thus we now switch our attention to a much more interesting and realistic model of shuffling.

Riffle shuffles This is what dealers do at the casino: They take the deck, split it into two parts, and these are then interleaved, for example by dropping cards from the bottoms of the two half-decks in some irregular pattern. Again a riffle shuffle performs a certain permutation on the cards in the deck, which we initially assume to be labelled from 1 to n, where 1 is the top card. The riffle shuffles correspond exactly to the permutations π ∈ Sn such that the sequence (π(1), π(2), . . . , π(n)) consists of two interlaced increasing sequences (only for the identity permutation it is one increasing sequence), and that there are exactly 2n − n distinct riffle shuffles on a deck of n cards. 0 0

1 2

1 1

3 4

1

5

0

1

0

2

1 1

3 4

1

5

0 1

1 3

1 0

4 2

1

5

In fact, if the pack is split such that the top t cards are taken into the right

hand  (0 ≤ t ≤ n) and the other n − t cards into the left hand, then there are n t ways to interleave the two hands, all of which generate distinct permutations — except that for each t there is one possibility to obtain the identity permutation. Now it’s not clear which probability distribution one should put on the riffle shuffles — there is no unique answer since amateurs and professional dealers would shuffle differently. However, the following model, developed first by Edgar N. Gilbert and Claude Shannon in 1955 (at the legendary

“A riffle shuffle”

192

Shuffling cards Bell Labs “Mathematics of Communication” department at the time), has several virtues: • it is elegant, simple, and seems natural, • it models quite well the way an amateur would perform riffle shuffles, • and we have a chance to analyze it. Here are three descriptions — all of them describe the same probability distribution Rif on Sn : 1. Rif : Sn −→ R is defined by ⎧ n+1 ⎪ if π = id, ⎨ 2n 1 Rif(π) := if π consists of two increasing sequences, 2n ⎪ ⎩ 0 otherwise.

 2. Cut off t cards from the deck with probability 21n nt , take them into your right hand, and take the rest of the deck into your left hand. Now when you have r cards in the right hand and  in the left, “drop” the r bottom card from your right hand with probability r+ , and from your  left hand with probability r+ . Repeat!

The inverse riffle shuffles correspond to the permutations π = (π(1), . . . , π(n)) that are increasing except for at most one “descent.” (Only the identity permutation has no descent.)

3. An inverse shuffle would take a subset of the cards in the deck, remove them from the deck, and place them on top of the remaining cards of the deck — while maintaining the relative order in both parts of the deck. Such a move is determined by the subset of the cards: Take all subsets with equal probability. Equivalently, assign a label “0” or “1” to each card, randomly and independently with probabilities 12 , and move the cards labelled “0” to the top. It is easy so see that these descriptions yield the same probability distributions. For (1) ⇐⇒ (3) just observe that we get the identity permutation whenever all the 0-cards are on top of all the cards that are assigned a 1. This defines the model. So how can we analyze it? How many riffle shuffles are needed to get close to random? We won’t get the precise best-possible answer, but quite a good one, by combining three components: (1) We analyze inverse riffle shuffles instead, (2) we describe a strong uniform stopping rule for these, (3) and show that the key to its analysis is given by the birthday paradox!

Theorem 2. After performing k riffle shuffles on a deck of n cards, the variation distance from a uniform distribution satisfies Rif

∗k

− U ≤ 1 −

n−1  i=1

i 1− k 2

 .

Shuffling cards

193

 Proof. (1) We may indeed analyze inverse riffle shuffles and try to see how fast they get us from the starting distribution to (close to) uniform. These inverse shuffles correspond to the probability distribution that is given by Rif(π) := Rif(π −1 ). Now the fact that every permutation has its unique inverse, and the fact that U(π) = U(π −1 ), yield Rif ∗k − U =  Rif

∗k

− U.

(This is Reeds’ inversion lemma!) (2) In every inverse riffle shuffle, each card gets associated a digit 0 or 1: 0 0

1 4

1 1

2 3

1

5

0 0

1 4

1 1

2 3

1

5

1 2 3 4 5

If we remember these digits — say we just write them onto the cards — then after k inverse riffle shuffles, each card has gotten an ordered string of k digits. Our stopping rule is: “STOP as soon as all cards have distinct strings.” When this happens, the cards in the deck are sorted according to the binary numbers bk bk−1 . . . b2 b1 , where bi is the bit that the card has picked up in the i-th inverse riffle shuffle. Since these bits are perfectly random and independent, this stopping rule is strong uniform! In the following example, for n = 5 cards, we need T = 3 inverse shuffles until we stop: 000 001

4 2

00 01

4 2

0 0

1 4

1 2

010 101 111

1 5 3

01 10 11

5 1 3

1 1 1

2 3 5

3 4 5

(3) The time T taken by this stopping rule is distributed according to the birthday paradox, for K = 2k : We put two cards into the same box if they have the same label bk bk−1 . . . b2 b1 ∈ {0, 1}k . So there are K = 2k boxes, and the probability that some box gets more than one card ist Prob[T > k] = 1 −

n−1 

1−

i=1

i 2k

 ,

and as we have seen this bounds the variation distance Rif ∗k − U = ∗k   Rif − U.

194

Shuffling cards

k 1 2 3 4 5 6 7 8 9 10

So how often do we have to shuffle? For large n we will need roughly k = 2 log2 (n) shuffles. Indeed, setting k := 2 log2 (cn) for some c ≥ 1 we 1 find (with a bit of calculus) that P [T > k] ≈ 1 − e− 2c2 ≈ 2c12 . Explicitly, for n = 52 cards the upper bound of Theorem 2 reads d(10) ≤ 0.73, d(12) ≤ 0.28, d(14) ≤ 0.08 — so k = 12 should be “random enough” for all practical purposes. But we don’t do 12 shuffles “in practice” — and they are not really necessary, as a more detailed analysis shows (with the results given in the margin). The analysis of riffle shuffles is part of a lively ongoing discussion about the right measure of what is “random enough.” Diaconis [4] is a guide to recent developments.

d(k) 1.000 1.000 1.000 1.000 0.952 0.614 0.334 0.167 0.085 0.043

The variation distance after k riffle shuffles, according to [2]

d(k) 1

Indeed, does it matter? Yes, it does: Even after three good riffle shuffles a sorted deck of 52 cards looks quite random . . . but it isn’t. Martin Gardner [5, Chapter 7] describes a number of striking card tricks that are based on the hidden order in such a deck!

References [1] D. A LDOUS & P. D IACONIS : Shuffling cards and stopping times, Amer. Math. Monthly 93 (1986), 333-348.

1

7

10

[2] D. BAYER & P. D IACONIS : Trailing the dovetail shuffle to its lair, Annals Applied Probability 2 (1992), 294-313. [3] E. B EHRENDS : Introduction to Markov Chains, Vieweg, Braunschweig/ Wiesbaden 2000. [4] P. D IACONIS : Mathematical developments from the analysis of riffle shuffling, in: “Groups, Combinatorics and Geometry. Durham 2001” (A. A. Ivanov, M. W. Liebeck and J. Saxl, eds.), World Scientific, Singapore 2003, pp. 73-97. [5] M. G ARDNER : Mathematical Magic Show, Knopf, New York/Allen & Unwin, London 1977. [6] E. N. G ILBERT: Theory of Shuffling, Technical Memorandum, Bell Laboratories, Murray Hill NJ, 1955.

Random enough?

Lattice paths and determinants

Chapter 29

The essence of mathematics is proving theorems — and so, that is what mathematicians do: They prove theorems. But to tell the truth, what they really want to prove, once in their lifetime, is a Lemma, like the one by Fatou in analysis, the Lemma of Gauss in number theory, or the Burnside– Frobenius Lemma in combinatorics. Now what makes a mathematical statement a true Lemma? First, it should be applicable to a wide variety of instances, even seemingly unrelated problems. Secondly, the statement should, once you have seen it, be completely obvious. The reaction of the reader might well be one of faint envy: Why haven’t I noticed this before? And thirdly, on an esthetic level, the Lemma — including its proof — should be beautiful! In this chapter we look at one such marvelous piece of mathematical reasoning, a counting lemma that first appeared in a paper by Bernt Lindström in 1972. Largely overlooked at the time, the result became an instant classic in 1985, when Ira Gessel and Gerard Viennot rediscovered it and demonstrated in a wonderful paper how the lemma could be successfully applied to a diversity of difficult combinatorial enumeration problems. The starting point is the usual permutation representation of the determinant of a matrix. Let M = (mij ) be a real n × n-matrix. Then  det M = sign σ m1σ(1) m2σ(2) · · · mnσ(n) , (1) σ

where σ runs through all permutations of {1, 2, . . . , n}, and the sign of σ is 1 or −1, depending on whether σ is the product of an even or an odd number of transpositions. Now we pass to graphs, more precisely to weighted directed bipartite graphs. A1 Ai An ... ... Let the vertices A1 , . . . , An stand for the rows of M , and B1 , . . . , Bn for mij the columns. For each pair of i and j draw an arrow from Ai to Bj and give m11 it the weight mij , as in the figure. mnj mi1 In terms of this graph, the formula (1) has the following interpretation: mnn ... ... • The left-hand side is the determinant of the path matrix M , whose B1 Bj Bn (i, j)-entry is the weight of the (unique) directed path from A to B . i

j

• The right-hand side is the weighted (signed) sum over all vertex-disjoint path systems from A = {A1 , . . . , An } to B = {B1 , . . . , Bn }. Such a system Pσ is given by paths A1 → Bσ(1) , . . . , An → Bσ(n) , M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_29, © Springer-Verlag Berlin Heidelberg 2013

196

Lattice paths and determinants and the weight of the path system Pσ is the product of the weights of the individual paths: w(Pσ ) = w(A1 → Bσ(1) ) · · · w(An → Bσ(n) ). In this interpretation formula (1) reads  det M = sign σ w(Pσ ). σ

An acyclic directed graph

And what is the result of Gessel and Viennot? It is the natural generalization of (1) from bipartite to arbitrary graphs. It is precisely this step which makes the Lemma so widely applicable — and what’s more, the proof is stupendously simple and elegant. Let us first collect the necessary concepts. We are given a finite acyclic directed graph G = (V, E), where acyclic means that there are no directed cycles in G. In particular, there are only finitely many directed paths between any two vertices A and B, where we include all trivial paths A → A of length 0. Every edge e carries a weight w(e). If P is a directed path from A to B, written shortly P : A → B, then we define the weight of P as  w(P ) := w(e), e∈P

which is defined to be w(P ) = 1 if P is a path of length 0. Now let A = {A1 , . . . , An } and B = {B1 , . . . , Bn } be two sets of n vertices, where A and B need not be disjoint. To A and B we associate the path matrix M = (mij ) with  mij := w(P ). P :Ai →Bj

A path system P from A to B consists of a permutation σ together with n paths Pi : Ai → Bσ(i) , for i = 1, . . . , n; we write sign P = sign σ . The weight of P is the product of the path weights w(P) =

n 

w(Pi ),

(2)

i=1

which is the product of the weights of all the edges of the path system. Finally, we say that the path system P = (P1 , . . . , Pn ) is vertex-disjoint if the paths of P are pairwise vertex-disjoint. Lemma. Let G = (V, E) be a finite weighted acyclic directed graph, A = {A1 , . . . , An } and B = {B1 , . . . , Bn } two n-sets of vertices, and M the path matrix from A to B. Then  det M = sign P w(P). (3) P vertex-disjoint path system

Lattice paths and determinants

197

 Proof. A typical summand of det(M ) is sign σ m1σ(1) · · · mnσ(n) , which can be written as  



 sign σ w(P1 ) · · · w(Pn ) . P1 :A1 →Bσ(1)

Pn :An →Bσ(n)

Summing over σ we immediately find from (2) that  sign P w(P), det M = P

where P runs through all path systems from A to B (vertex-disjoint or not). Hence to arrive at (3), all we have to show is  sign P w(P) = 0 , (4) P∈N

where N is the set of all path systems that are not vertex-disjoint. And this is accomplished by an argument of singular beauty. Namely, we exhibit an involution π : N → N (without fixed points) such that for P and πP w(πP) = w(P)

and

sign πP

= −sign P .

Clearly, this will imply (4) and thus the formula (3) of the Lemma. The involution π is defined in the most natural way. Let P ∈ N with paths Pi : Ai → Bσ(i) . By definition, some pair of paths will intersect: • Let i0 be the minimal index such that Pi0 shares some vertex with another path. • Let X be the first such common vertex on the path Pi0 .

Ai0

Aj0

X

• Let j0 be the minimal index (j0 > i0 ) such that Pj0 has the vertex X in common with Pi0 . Now we construct the new system πP = (P1 , . . . , Pn ) as follows: • Set Pk = Pk for all k = i0 , j0 . • The new path Pi 0 goes from Ai0 to X along Pi0 , and then continues to Bσ(j0 ) along Pj0 . Similarly, Pj 0 goes from Aj0 to X along Pj0 and continues to Bσ(i0 ) along Pi0 . Clearly π(πP) = P, since the index i0 , the vertex X, and the index j0 are the same as before. In other words, applying π twice we switch back to the old paths Pi . Next, since πP and P use precisely the same edges, we certainly have w(πP) = w(P). And finally, since the new permutation σ is obtained by multiplying σ with the transposition (i0 , j0 ), we find that sign πP = −sign P , and that’s it.  The Gessel–Viennot Lemma can be used to derive all basic properties of determinants, just by looking at appropriate graphs. Let us consider one particularly striking example, the formula of Binet–Cauchy, which gives a very useful generalization of the product rule for determinants.

Bσ(j0 )

Bσ(i0 )

198

Lattice paths and determinants

A1

...

Ai

...

Z

where PZ is the (r × r)-submatrix of P with column-set Z, and QZ the (r × r)-submatrix of Q with the corresponding rows Z.

Bk

...

Bs qkj

q1j ... C1

Ar

pik

pi1 B1

Theorem. If P is an (r × s)-matrix and Q an (s × r)-matrix, r ≤ s, then  (det PZ )(det QZ ), det(P Q) =

... Cj

Cr

 Proof. Let the bipartite graph on A and B correspond to P as before, and similarly the bipartite graph on B and C to Q. Consider now the concatenated graph as indicated in the figure on the left, and observe that

the (i, j)entry mij of the path matrix M from A to C is precisely mij = k pik qkj , thus M = P Q. Since the vertex-disjoint path systems from A to C in the concatenated graph correspond to pairs of systems from A to Z resp. from Z to C, the result follows immediately from the Lemma, by noting that sign (στ ) = (sign σ) (sign τ ).  The Lemma of Gessel–Viennot is also the source of a great number of results that relate determinants to enumerative properties. The recipe is always the same: Interpret the matrix M as a path matrix, and try to compute the right-hand side of (3). As an illustration we will consider the original problem studied by Gessel and Viennot, which led them to their Lemma: Suppose that a1 < a2 < . . . < an and b1 < b2 < . . . < bn are two sets of natural numbers. We wish to compute the determinant

of the matrix M = (mij ), where mij is the binomial coefficient abji .

1 1 1 1 1 1 1 1

1 2 3 4 5 6 7

1 3 1 6 4 1 10 10 5 1 15 20 15 6 21 35 35 21

a=3

b=4

1 7

In other words, Gessel and Viennot were looking at the determinants of arbitrary square matrices of Pascal’s triangle, such as the matrix ⎛ 3 3 3 ⎞ ⎞ ⎛ 1 3 4 3 1 0 ⎜ 4 4 4 ⎟ ⎟ = det ⎝ 4 4 1 ⎠ det ⎜ 3 4 ⎠ ⎝ 1

6 6 6 6 20 15 1

1

3

4

given by the bold entries of Pascal’s triangle, as displayed in the margin. As a preliminary step to the solution of the problem we recall a well-known result which connects binomial coefficients to lattice paths. Consider an a × b-lattice as in the margin. Then the number of paths from the lower left-hand corner to the upper right-hand corner, where the only steps

that are allowed for the paths are up (North) and to the right (East), is a+b a . The proof of this is easy: each path consists of an arbitrary sequence of b “east” and a “north” steps, and thus it can be encoded by a sequence of the form NENEEEN, consisting of a+b letters, a N’s and b E’s. The number of such strings is the number of ways to choose a positions of letters N from  a+b  a total of a + b positions, which is a+b = . a b

Lattice paths and determinants

199

Now look at the figure to the right, where Ai is placed at the point (0, −ai ) and Bj at (bj , −bj ). The number of paths from Ai to Bj in this grid that use only steps to the



 north and east is, by what we just proved, bj +(abji −bj ) = abji . In other words, the matrix of binomials M is precisely the path matrix from A to B in the directed lattice graph for which all edges have weight 1, and all edges are directed to go north or east. Hence to compute det M we may apply the Gessel–Viennot Lemma. A moment’s thought shows that every vertexdisjoint path system P from A to B must consist of paths Pi : Ai → Bi for all i. Thus the only possible permutation is the identity, which has sign = 1, and we obtain the beautiful result   det abji = # vertex-disjoint path systems from A to B. In particular, this implies the far from obvious fact that det M is always nonnegative, since the right-hand side of the equality counts something. More precisely, one gets from the Gessel–Viennot Lemma that det M = 0 if and only if ai < bi for some i. In our previous small example,

B1 A1 .. .

Ai Bn .. .

An

1

⎛ 3 3 3 ⎞ 1

3

4

⎜ 4 4 4 ⎟ ⎟ = # vertex-disjoint det ⎜ 3 4 ⎠ ⎝ 1 path systems in

6 6 6 1

3

3 4

Bj

3 4

4 6

“Lattice paths”

200

Lattice paths and determinants

References [1] I. M. G ESSEL & G. V IENNOT: Binomial determinants, paths, and hook length formulae, Advances in Math. 58 (1985), 300-321. [2] B. L INDSTRÖM : On the vector representation of induced matroids, Bulletin London Math. Soc. 5 (1973), 85-90.

Cayley’s formula for the number of trees

Chapter 30

One of the most beautiful formulas in enumerative combinatorics concerns the number of labeled trees. Consider the set N = {1, 2, . . . , n}. How many different trees can we form on this vertex set? Let us denote this number by Tn . Enumeration “by hand” yields T1 = 1, T2 = 1, T3 = 3, T4 = 16, with the trees shown in the following table: 1

1

2

1

2

1

2

1

2

1

2

1

2

1

2

3 1

2

3 1

2

1

3 2

1

2

1

2

3

4

3

4

3

4

3

4

3

4

3

4

3

4

3

4

1

2

1

2

1

2

1

2

1

2

1

2

1

2

1

2

3

4

3

4

3

4

3

4

3

4

3

4

3

4

3

4

Note that we consider labeled trees, that is, although there is only one tree of order 3 in the sense of graph isomorphism, there are 3 different labeled trees obtained by marking the inner vertex 1, 2 or 3. For n = 5 there are three non-isomorphic trees:

5

60

60

For the first tree there are clearly 5 different labelings, and for the second and third there are 5! 2 = 60 labelings, so we obtain T5 = 125. This should be enough to conjecture Tn = nn−2 , and that is precisely Cayley’s result. Theorem. There are nn−2 different labeled trees on n vertices. This beautiful formula yields to equally beautiful proofs, drawing on a variety of combinatorial and algebraic techniques. We will outline three of them before presenting the proof which is to date the most beautiful of them all. M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_30, © Springer-Verlag Berlin Heidelberg 2013

Arthur Cayley

202

Cayley’s formula for the number of trees

1 2

1

1

2

2

 First proof (Bijection). The classical and most direct method is to find a bijection from the set of all trees on n vertices onto another set whose cardinality is known to be nn−2 . Naturally, the set of all ordered sequences (a1 , . . . , an−2 ) with 1 ≤ ai ≤ n comes into mind. Thus we want to uniquely encode every tree T by a sequence (a1 . . . , an−2 ). Such a code was found by Prüfer and is contained in most books on graph theory. Here we want to discuss another bijection proof, due to Joyal, which is less known but of equal elegance and simplicity. For this, we consider not just trees t on N = {1, . . . , n} but trees together with two distinguished vertices, the left end and the right end , which may coincide. Let Tn = {(t; , )} be this new set; then, clearly, |Tn | = n2 Tn .

1 2

The four trees of T2

f G

9

4

7

1

10 3

5

8

is represented by the directed graph in the margin.  f . Since there is precisely one edge emanating Look at a component of G from each vertex, the component contains equally many vertices and edges, and hence precisely one directed cycle. Let M ⊆ N be the union of the vertex sets of these cycles. A moment’s thought shows that M is the unique maximal subset of N such  M acts as a bijection  that the restriction of f onto a b ... z on M . Write f |M = such that the numbers f (a) f (b) . . . f (z) a, b, . . . , z in the first row appear in natural order. This gives us an ordering f (a), f (b), . . . , f (z) of M according to the second row. Now f (a) is our left end and f (z) is our right end.

6

2

Our goal is thus to prove |Tn | = nn . Now there is a set whose size is known to be nn , namely the set N N of all mappings from N into N . Thus our formula is proved if we can find a bijection from N N onto Tn .  f by Let f : N −→ N be any map. We represent f as a directed graph G drawing arrows from i to f (i). For example, the map   1 2 3 4 5 6 7 8 9 10 f= 7 5 5 9 1 2 5 8 4 7

The tree t corresponding to the map f is now constructed as follows: Draw f (a), . . . , f (z) in this order as a path from f (a) to f (z), and fill in the  f (deleting the arrows). remaining vertices as in G 9

1

5

4

7 10

In our example above we obtain M = {1, 4, 5, 7, 8, 9}

8

2

3

6

 f |M =

1 4 7 9

5 7 1 5

8 9 8 4



and thus the tree t depicted in the margin. It is immediate how to reverse this correspondence: Given a tree t, we look at the unique path P from the left end to the right end. This gives us the set M and the mapping f |M . The remaining correspondences i → f (i) are then filled in according to the unique paths from i to P . 

Cayley’s formula for the number of trees

203

 Second proof (Linear Algebra). We can think of Tn as the number of spanning trees in the complete graph Kn . Now let us look at an arbitrary connected simple graph G on V = {1, 2, . . . , n}, denoting by t(G) the number of spanning trees; thus Tn = t(Kn ). The following celebrated result is Kirchhoff’s matrix-tree theorem (see [1]). Consider the incidence matrix B = (bie ) of G, whose rows are labeled by V , the columns by E, where we write bie = 1 or 0 depending on whether i ∈ e or i ∈ e. Note that |E| ≥ n − 1 since G is connected. In every column we replace one of the two 1’s by −1 in an arbitrary manner (this amounts to an orientation of G), and call the new matrix C. M = CC T is then a symmetric (n × n)-matrix with the degrees d1 , . . . , dn in the main diagonal. Proposition. We have t(G) = det Mii for all i = 1, . . . , n, where Mii results from M by deleting the i-th row and the i-th column.  Proof. The key to the proof is the Binet–Cauchy theorem proved in the previous chapter: When P is an (r × s)-matrix and Q an (s × r)-matrix, r ≤ s, then det(P Q) equals the sum of the products of determinants of corresponding (r × r)-submatrices, where “corresponding” means that we take the same indices for the r columns of P and the r rows of Q. For Mii this means that   det N · det N T = (det N )2 , det Mii = N

N

where N runs through all (n − 1) × (n − 1) submatrices of C\{row i}. The n − 1 columns of N correspond to a subgraph of G with n − 1 edges on n vertices, and it remains to show that  ±1 if these edges span a tree det N = 0 otherwise. Suppose the n − 1 edges do not span a tree. Then there exists a component which does not contain i. Since the corresponding rows of this component add to 0, we infer that they are linearly dependent, and hence det N = 0. Assume now that the columns of N span a tree. Then there is a vertex j1 = i of degree 1; let e1 be the incident edge. Deleting j1 , e1 we obtain a tree with n − 2 edges. Again there is a vertex j2 = i of degree 1 with incident edge e2 . Continue in this way until j1 , j2 , . . . , jn−1 and e1 , e2 , . . . , en−1 with ji ∈ ei are determined. Now permute the rows and columns to bring jk into the k-th row and ek into the k-th column. Since by construction jk ∈ / e for k < , we see that the new matrix N is lower triangular with all elements on the main diagonal equal to ±1. Thus det N = ±det N = ±1, and we are done. For the special case G = Kn we clearly obtain ⎛ ⎞ n−1 −1 ... −1 ⎜ −1 n−1 −1 ⎟ ⎟ .. .. Mii = ⎜ .. ⎝ ⎠ . . . −1 −1 ... n − 1 and an easy computation shows det Mii = nn−2 .



“A nonstandard method of counting trees: Put a cat into each tree, walk your dog, and count how often he barks.”

204

Cayley’s formula for the number of trees  Third proof (Recursion). Another classical method in enumerative combinatorics is to establish a recurrence relation and to solve it by induction. The following idea is essentially due to Riordan and Rényi. To find the proper recursion, we consider a more general problem (which already appears in Cayley’s paper). Let A be an arbitrary k-set of the vertices. By Tn,k we denote the number of (labeled) forests on {1, . . . , n} consisting of k trees where the vertices of A appear in different trees. Clearly, the set A does not matter, only the size k. Note that Tn,1 = Tn .

For example, T4,2 = 8 for A = {1, 2}

1 2 3

"

#$ i

...

k

1

2

1

2

1

2

1

2

1

2

1

2

1

2

1

2

3

4

3

4

3

4

3

4

3

4

3

4

3

4

3

4

Consider such a forest F with A = {1, 2, . . . , k}, and suppose 1 is adjacent to i vertices, as indicated in the margin. Deleting 1, the i neighbors together with 2, . . . , k yield one vertex each in the components of a forest that consists of k − 1 + i trees. As we can (re)construct F by first fixing i, then choosing the i neighbors of 1 and then the forest F \1, this yields Tn,k =

%

n−k  i=0

 n−k Tn−1,k−1+i i

(1)

for all n ≥ k ≥ 1, where we set T0,0 = 1, Tn,0 = 0 for n > 0. Note that T0,0 = 1 is necessary to ensure Tn,n = 1. Proposition. We have and thus, in particular,

Tn,k = k nn−k−1

(2)

Tn,1 = Tn = nn−2 .  Proof. By (1), and using induction, we find Tn,k =

=

=

=

=

n−k 

 n−k (k − 1 + i)(n − 1)n−1−k−i (i → n − k − i) i i=0 n−k  n − k  (n − 1 − i)(n − 1)i−1 i i=0 n−k n−k  n − k   n − k  i (n − 1) − i(n − 1)i−1 i i i=0 i=1 n−k  n − 1 − k  (n − 1)i−1 nn−k − (n − k) i − 1 i=1 n−1−k  n − 1 − k  (n − 1)i nn−k − (n − k) i i=0

= nn−k − (n − k)nn−1−k = k nn−1−k .



Cayley’s formula for the number of trees

205

 Fourth proof (Double Counting). The following marvelous idea due to Jim Pitman gives Cayley’s formula and its generalization (2) without induction or bijection — it is just clever counting in two ways. A rooted forest on {1, . . . , n} is a forest together with a choice of a root in each component tree. Let Fn,k be the set of all rooted forests that consist of k rooted trees. Thus Fn,1 is the set of all rooted trees. Note that |Fn,1 | = nTn , since in every tree there are n choices for the root. We now regard Fn,k ∈ Fn,k as a directed graph with all edges directed away from the roots. Say that a forest F contains another forest F if F contains F as directed graph. Clearly, if F properly contains F , then F has fewer components than F . The figure shows two such forests with the roots on top. Here is the crucial idea. Call a sequence F1 , . . . , Fk of forests a refining sequence if Fi ∈ Fn,i and Fi contains Fi+1 , for all i. Now let Fk be a fixed forest in Fn,k and denote

F2

4

8 3 7

6 9

1 10 F2 contains F3

8 3

• by N ∗ (Fk ) the number of refining sequences ending in Fk .

2

We count N ∗ (Fk ) in two ways, first by starting at a tree and secondly by starting at Fk . Suppose F1 ∈ Fn,1 contains Fk . Since we may delete the k − 1 edges of F1 \Fk in any possible order to get a refining sequence from F1 to Fk , we find

F3

(4)

and out comes, with (3), the unexpectedly simple relation N (Fk ) = nk−1

5

9

1

6

10

(3)

Let us now start at the other end. To produce from Fk an Fk−1 we have to add a directed edge, from any vertex a, to any of the k − 1 roots of the trees that do not contain a (see the figure on the right, where we pass from F3 to F2 by adding the edge 3 7). Thus we have n(k − 1) choices. Similarly, for Fk−1 we may produce a directed edge from any vertex b to any of the k−2 roots of the trees not containing b. For this we have n(k−2) choices. Continuing this way, we arrive at N ∗ (Fk ) = nk−1 (k − 1)!,

7

4

• by N (Fk ) the number of rooted trees containing Fk , and

N ∗ (Fk ) = N (Fk ) (k − 1)!.

5

2

for any Fk ∈ Fn,k .

For k = n, Fn consists just of n isolated vertices. Hence N (Fn ) counts the number of all rooted trees, and we obtain |Fn,1 | = nn−1 , and thus Cayley’s formula.  But we get even more out of this proof. Formula (4) yields for k = n: 7 8 # refining sequences (F1 , F2 , . . . , Fn ) = nn−1 (n − 1)!. (5) For Fk ∈Fn,k , let N ∗∗ (Fk ) denote the number of those refining sequences F1 , . . . , Fn whose k-th term is Fk . Clearly this is N ∗ (Fk ) times the number

8

2

7

4

3

F3 −→ F2

5 6

9

1 10

206

Cayley’s formula for the number of trees of ways to choose (Fk+1 , . . . , Fn ). But this latter number is (n − k)! since we may delete the n − k edges of Fk in any possible way, and so N ∗∗ (Fk ) = N ∗ (Fk )(n − k)! = nk−1 (k − 1)!(n − k)!.

(6)

Since this number does not depend on the choice of Fk , dividing (5) by (6) yields the number of rooted forests with k trees:   n nn−1 (n − 1)! = k nn−1−k . |Fn,k | = k−1 n (k − 1)!(n − k)! k

 As we may choose the k roots in nk possible ways, we have reproved the formula Tn,k = knn−k−1 without recourse to induction. Let us end with a historical note. Cayley’s paper from 1889 was anticipated by Carl W. Borchardt (1860), and this fact was acknowledged by Cayley himself. An equivalent result appeared even earlier in a paper of James J. Sylvester (1857), see [2, Chapter 3]. The novelty in Cayley’s paper was the use of graph theory terms, and the theorem has been associated with his name ever since.

References [1] M. A IGNER : Combinatorial Theory, Springer–Verlag, Berlin Heidelberg New York 1979; Reprint 1997. [2] N. L. B IGGS , E. K. L LOYD & R. J. W ILSON : Graph Theory 1736-1936, Clarendon Press, Oxford 1976. [3] A. C AYLEY: A theorem on trees, Quart. J. Pure Appl. Math. 23 (1889), 376-378; Collected Mathematical Papers Vol. 13, Cambridge University Press 1897, 26-28. [4] A. J OYAL : Une théorie combinatoire des séries formelles, Advances in Math. 42 (1981), 1-82. [5] J. P ITMAN : Coalescent random forests, J. Combinatorial Theory, Ser. A 85 (1999), 165-193. [6] H. P RÜFER : Neuer Beweis eines Satzes über Permutationen, Archiv der Math. u. Physik (3) 27 (1918), 142-144. [7] A. R ÉNYI : Some remarks on the theory of trees. MTA Mat. Kut. Inst. Kozl. (Publ. math. Inst. Hungar. Acad. Sci.) 4 (1959), 73-85; Selected Papers Vol. 2, Akadémiai Kiadó, Budapest 1976, 363-374. [8] J. R IORDAN : Forests of labeled trees, J. Combinatorial Theory 5 (1968), 90-103.

Identities versus bijections

Chapter 31

Consider the infinite product (1 + x)(1

+ x2 )(1 + x3 )(1 + x4 ) · · · and expand it in the usual way into a series n≥0 an xn by grouping together those products that yield the same power xn . By inspection we find for the first terms  (1 + xk ) = 1 + x + x2 + 2x3 + 2x4 + 3x5 + 4x6 + 5x7 + . . . . (1) k≥1

So we have e. g. a6 = 4, a7 = 5, and we (rightfully) suspect that an goes to infinity with n −→ ∞. Looking at the equally simple product (1 − x)(1 − x2 )(1 − x3 )(1 − x4 ) · · · something unexpected happens. Expanding this product we obtain  (1 − xk ) = 1 − x − x2 + x5 + x7 − x12 − x15 + x22 + x26 − . . . . (2) k≥1

It seems that all coefficients are equal to 1, −1 or 0. But is this true? And if so, what is the pattern? Infinite sums and products and their convergence have played a central role in analysis since the invention of the calculus, and contributions to the subject have been made by some of the greatest names in the field, from Leonhard Euler to Srinivasa Ramanujan. In explaining identities such as (1) and (2), however, we disregard convergence questions — we simply manipulate the coefficients. In the language of the trade we deal with “formal” power series and products. In this framework we are going to show how combinatorial arguments lead to elegant proofs of seemingly difficult identities. Our basic notion is that of a partition of a natural number. We call any sum λ : n = λ1 + λ2 + . . . + λt

with λ1 ≥ λ2 ≥ · · · ≥ λt ≥ 1

a partition of n. P (n) shall be the set of all partitions of n, with p(n) := |P (n)|, where we set p(0) = 1. What have partitions got to do with our problem? Well, consider the following product of infinitely many series: (1+x+x2 +x3 +. . . )(1+x2 +x4 +x6 +. . . )(1+x3 +x6 +x9 +. . .) · · · (3) where the k-th factor is (1 + xk + x2k + x3k + . .

. ). What is the coefficient of xn when we expand this product into a series n≥0 an xn ? A moment’s M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_31, © Springer-Verlag Berlin Heidelberg 2013

5=5 5 = 4+1 5 = 3+2 5 = 3+1+1 5 = 2+2+1 5 = 2+1+1+1 5 = 1 + 1 + 1 + 1 + 1. The partitions counted by p(5) = 7

208

Identities versus bijections thought should convince you that this is just the number of ways to write n as a sum n

=

n1 · 1 + n2 · 2 + n3 · 3 + . . .

=

1 + ...+ 1 + 2 + ... + 2 + 3 + ...+ 3 + ... . " #$ % " #$ % " #$ % n1 n2 n3

So the coefficient is nothing else but the number p(n) of partitions of n. 1 Since the geometric series 1 + xk + x2k + . . . equals 1−x k , we have proved our first identity:   1 = p(n) xn . (4) 1 − xk k≥1

n≥0

1 What’s more, we see from our analysis that the factor 1−x k accounts for 1 the contribution of k to a partition of n. Thus, if we leave out 1−x k from the product on the left side of (4), then k does not appear in any partition on the right side. As an example we immediately obtain

6=5+1 6=3+3 6 = 3+1+1+1 6 = 1+1+1+1+1+1 Partitions of 6 into odd parts: po (6) = 4

 i≥1

 1 = po (n) xn , 2i−1 1−x

where po (n) is the number of partitions of n all of whose summands are odd, and the analogous statement holds when all summands are even. By now it should be clear what the n-th coefficient in the infinite product  k k k≥1 (1 + x ) will be. Since we take from any factor in (3) either 1 or x , this means that we consider only those partitions where any summand k appears at most once. In other words, our original product (1) is expanded into   (1 + xk ) = pd (n) xn , (6) k≥1

7=7 7 =5+1+1 7 =3+3+1 7 = 3+1+1+1+1 7 = 1+1+1+1+1+1+1 7=7 7=6+1 7=5+2 7=4+3 7 = 4 + 2 + 1. The partitions of 7 into odd resp. distinct parts: po (7) = pd (7) = 5.

(5)

n≥0

n≥0

where pd (n) is the number of partitions of n into distinct summands. Now the method of formal series displays its full power. Since 1 − x2 = (1 − x)(1 + x) we may write  k≥1

(1 + xk ) =

 1 − x2k  1 = k 1−x 1 − x2k−1

k≥1

k≥1

since all factors 1 − x2i with even exponent cancel out. So, the infinite products in (5) and (6) are the same, and hence also the series, and we obtain the beautiful result po (n) = pd (n)

for all n ≥ 0.

(7)

Such a striking equality demands a simple proof by bijection — at least that is the point of view of any combinatorialist.

Identities versus bijections

209

Problem. Let Po (n) and Pd (n) be the partitions of n into odd and into distinct summands, respectively: Find a bijection from Po (n) onto Pd (n)! Several bijections are known, but the following one due to J. W. L. Glaisher (1907) is perhaps the neatest. Let λ be a partition of n into odd parts. We collect equal summands and have n

= =

λ1 + . . . + λ1 + λ2 + . . . + λ2 + . . . + λt + . . . + λt " #$ % " #$ % " #$ % n1 n2 nt n1 · λ1 + n2 · λ2 + . . . + nt · λt .

Now we write n1 = 2m1 + 2m2 + . . . + 2mr in its binary representation and similarly for the other ni . The new partition λ of n is then λ : n = 2m1 λ1 + 2m2 λ1 + . . . + 2mr λ1 + 2k1 λ2 + . . . . We have to check that λ is in Pd (n), and that φ : λ −→ λ is indeed a bijection. Both claims are easy to verify: If 2a λi = 2b λj then 2a = 2b since λi and λj are odd, and so λi = λj . Hence λ is in Pd (n). Conversely, when n = μ1 + μ2 + . . . + μs is a partition into distinct summands, then we reverse the bijection by collecting all μi with the same highest power of 2, and write down the odd parts with the proper multiplicity. The margin displays an example. Manipulating formal products has thus led to the equality po (n) = pd (n) for partitions which we then verified via a bijection. Now we turn this around, give a bijection proof for partitions and deduce an identity. This time our goal is to identify the pattern in the expansion (2). Look at

For example, λ : 25 = 5+5+5+3+3+1+1+1+1 is mapped by φ to λ : 25 = (2+1)5 + (2)3 + (4)1 = 10 + 5 + 6 + 4 = 10 + 6 + 5 + 4 .

We write λ : 30 = 12 + 6 + 5 + 4 + 3 as 30 = 4(3+1) + 2(3) + 1(5+3) = (1)5 + (4+2+1)3 + (4)1 and obtain as φ−1 (λ ) the partition λ : 30 = 5 + 3 + 3 + 3 + 3 + 3 + 3 + 3+1+1+1+1 into odd summands.

4

1 − x − x2 + x5 + x7 − x12 − x15 + x22 + x26 − x35 − x40 + . . . . The exponents (apart from 0) seem to come in pairs, and taking the exponents of the first power in each pair gives the sequence 1 5

12 22 35 51 70

k≥1

(1 − xk ) = 1 +

 j≥1

2 j=1

...

well-known to Euler. These are the pentagonal numbers f (j), whose name is suggested by the figure in the margin. 2 2 We easily compute f (j) = 3j 2−j and f¯(j) = 3j 2+j for the other number of each pair. In summary, we conjecture, as Euler has done, that the following formula should hold. Theorem. 

3

 3j2 −j  3j 2 +j . (−1)j x 2 + x 2

(8)

Pentagonal numbers

210

Identities versus bijections Euler proved this remarkable theorem by calculations with formal series, but we give a bijection proof from The Book. First of all, we notice by (4)  that the product k≥1 (1−xk ) is precisely the inverse of our partition series



n k n n≥0 p(n)x . Hence setting k≥1 (1 − x ) =: n≥0 c(n)x , we find

    c(n)xn · p(n)xn = 1. n≥0

n≥0

Comparing coefficients this means that c(n) is the unique sequence with c(0) = 1 and n 

c(k)p(n − k) = 0

k=0

Writing the right-hand of (8) as

∞

for all n ≥ 1.

(−1)j x

3j 2 +j 2

(9)

, we have to show that

j=−∞

⎧ 2 ⎪ 1 for k = 3j 2+j , when j ∈ Z is even, ⎪ ⎨ 2 c(k) = −1 for k = 3j 2+j , when j ∈ Z is odd, ⎪ ⎪ ⎩ 0 otherwise 2

gives this unique sequence. Setting b(j) = 3j 2+j for j ∈ Z and substituting these values into (9), our conjecture takes on the simple form   p(n − b(j)) = p(n − b(j)) for all n, j even

j odd

where of course we only consider j with b(j) ≤ n. So the stage is set: We have to find a bijection   φ: P (n − b(j)) −→ P (n − b(j)). j even

As an example consider n = 15, j = 2, so b(2) = 7. The partition 3 + 2 + 2 + 1 in P (15 − b(2)) = P (8) is mapped to 9+2+1+1, which is in P (15−b(1)) = P (13).

j odd

Again several bijections have been suggested, but the following construction by David Bressoud and Doron Zeilberger is astonishingly simple. We just give the definition of φ (which is, in fact, an involution), and invite the reader to verify the easy details. For λ : λ1 + . . . + λt ∈ P (n − b(j)) set ⎧ ⎪ ⎨(t + 3j − 1) + (λ1 − 1) + . . . + (λt − 1) if t + 3j ≥ λ1 , φ(λ) := ⎪ ⎩(λ2 + 1) + . . . + (λt + 1) + 1 + . . . + 1 if t + 3j < λ1 , " #$ % λ1 −t−3j−1

where we leave out possible 0’s. One finds that in the first case φ(λ) is in P (n − b(j − 1)), and in the second case in P (n − b(j + 1)). This was beautiful, and we can get even more out of it. We already know that   (1 + xk ) = pd (n) xn . k≥1

n≥0

Identities versus bijections

211

As experienced formal series manipulators we notice that the introduction of the new variable y yields   (1 + yxk ) = pd,m (n) xn y m , k≥1

n,m≥0

where pd,m (n) counts the partitions of n into precisely m distinct summands. With y = −1 this yields   (1 − xk ) = (Ed (n) − Od (n)) xn , (10) k≥1

n≥0

where Ed (n) is the number of partitions of n into an even number of distinct parts, and Od (n) is the number of partitions into an odd number. And here is the punchline. Comparing (10) to Euler’s expansion in (8) we infer the beautiful result ⎧ 2 ⎪ 1 for n = 3j 2±j when j ≥ 0 is even, ⎪ ⎪ ⎨ 2 Ed (n) − Od (n) = −1 for n = 3j 2±j when j ≥ 1 is odd, ⎪ ⎪ ⎪ ⎩ 0 otherwise.

An example for n = 10: 10 = 9 + 1 10 = 8 + 2 10 = 7 + 3 10 = 6 + 4 10 = 4 + 3 + 2 + 1 and 10 = 10 10 = 7 + 2 + 1 10 = 6 + 3 + 1 10 = 5 + 4 + 1 10 = 5 + 3 + 2, so Ed (10) = Od (10) = 5.

This is, of course, just the beginning of a longer and still ongoing story. The theory of infinite products is replete with unexpected indentities, and with their bijective counterparts. The most famous examples are the so-called Rogers–Ramanujan identities, named after Leonard Rogers and Srinivasa Ramanujan, in which the number 5 plays a mysterious role:  k≥1

 k≥1

1 (1 − x5k−4 )(1 − x5k−1 ) 1 (1 − x5k−3 )(1 − x5k−2 )

=

 n≥0

=

 n≥0

2

xn , (1 − x)(1 − x2 ) · · · (1 − xn ) 2

xn +n . (1 − x)(1 − x2 ) · · · (1 − xn )

The reader is invited to translate them into the following partition identities first noted by Percy MacMahon: • Let f (n) be the number of partitions of n all of whose summands are of the form 5k + 1 or 5k + 4, and g(n) the number of partitions whose summands differ by at least 2. Then f (n) = g(n). • Let r(n) be the number of partitions of n all of whose summands are of the form 5k + 2 or 5k + 3, and s(n) the number of partitions whose parts differ by at least 2 and which do not contain 1. Then r(n) = s(n). All known formal series proofs of the Rogers–Ramanujan identities are quite involved, and for a long time bijection proofs of f (n) = g(n) and of r(n) = s(n) seemed elusive. Such proofs were eventually given 1981 by Adriano Garsia and Stephen Milne. Their bijections are, however, very complicated — Book proofs are not yet in sight.

Srinivasa Ramanujan

212

Identities versus bijections

References [1] G. E. A NDREWS : The Theory of Partitions, Encyclopedia of Mathematics and its Applications, Vol. 2, Addison-Wesley, Reading MA 1976. [2] D. B RESSOUD & D. Z EILBERGER : Bijecting Euler’s partitions-recurrence, Amer. Math. Monthly 92 (1985), 54-55. [3] A. G ARSIA & S. M ILNE : A Rogers–Ramanujan bijection, J. Combinatorial Theory, Ser. A 31 (1981), 289-339. [4] S. R AMANUJAN : Proof of certain identities in combinatory analysis, Proc. Cambridge Phil. Soc. 19 (1919), 214-216. [5] L. J. ROGERS : Second memoir on the expansion of certain infinite products, Proc. London Math. Soc. 25 (1894), 318-343.

Completing Latin squares

Some of the oldest combinatorial objects, whose study apparently goes back to ancient times, are the Latin squares. To obtain a Latin square, one has to fill the n2 cells of an (n × n)-square array with the numbers 1, 2, . . . , n so that that every number appears exactly once in every row and in every column. In other words, the rows and columns each represent permutations of the set {1, . . . , n}. Let us call n the order of the Latin square. Here is the problem we want to discuss. Suppose someone started filling the cells with the numbers {1, 2, . . . , n}. At some point he stops and asks us to fill in the remaining cells so that we get a Latin square. When is this possible? In order to have a chance at all we must, of course, assume that at the start of our task any element appears at most once in every row and in every column. Let us give this situation a name. We speak of a partial Latin square of order n if some cells of an (n × n)-array are filled with numbers from the set {1, . . . , n} such that every number appears at most once in every row and column. So the problem is: When can a partial Latin square be completed to a Latin square of the same order? Let us look at a few examples. Suppose the first n − 1 rows are filled and the last row is empty. Then we can easily fill in the last row. Just note that every element appears n − 1 times in the partial Latin square and hence is missing from exactly one column. Hence by writing each element below the column where it is missing we have completed the square correctly. Going to the other end, suppose only the first row is filled. Then it is again easy to complete the square by cyclically rotating the elements one step in each of the following rows. So, while in our first example the completion is forced, we have lots of possibilities in the second example. In general, the fewer cells are prefilled, the more freedom we should have in completing the square. However, the margin displays an example of a partial square with only n cells filled which clearly cannot be completed, since there is no way to fill the upper right-hand corner without violating the row or column condition.

If fewer than n cells are filled in an (n × n)-array, can one then always complete it to obtain a Latin square?

M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_32, © Springer-Verlag Berlin Heidelberg 2013

Chapter 32

1

2

3

4

2

1

4

3

4

3

1

2

3

4

2

1

A Latin square of order 4

1

4

2

5

3

4

2

5

3

1

2

5

3

1

4

5

3

1

4

2

3

1

4

2

5

A cyclic Latin square

1

2

...

n-1 n

A partial Latin square that cannot be completed

214

Completing Latin squares This question was raised by Trevor Evans in 1960, and the assertion that a completion is always possible quickly became known as the Evans conjecture. Of course, one would try induction, and this is what finally led to success. But Bohdan Smetaniuk’s proof from 1981, which answered the question, is a beautiful example of just how subtle an induction proof may be needed in order to do such a job. And, what’s more, the proof is constructive, it allows us to complete the Latin square explicitly from any initial partial configuration.

1

3

2

2

1

3

3

2

1

R: 1 1 1 2 2 2 3 3 3 C: 1 2 3 1 2 3 1 2 3 E: 1 3 2 2 1 3 3 2 1

If we permute the lines of the above example cyclically, R −→ C −→ E −→ R, then we obtain the following line array and Latin square:

1

2

3

3

1

2

2

3

1

R: 1 3 2 2 1 3 3 2 1 C: 1 1 1 2 2 2 3 3 3 E: 1 2 3 1 2 3 1 2 3

Before proceeding to the proof let us take a closer look at Latin squares in general. We can alternatively view a Latin square as a (3 × n2 )-array, called the line array of the Latin square. The figure to the left shows a Latin square of order 3 and its associated line array, where R, C and E stand for rows, columns and elements. The condition on the Latin square is equivalent to saying that in any two lines of the line array all n2 ordered pairs appear (and therefore each pair appears exactly once). Clearly, we may permute the symbols in each line arbitrarily (corresponding to permutations of rows, columns or elements) and still obtain a Latin square. But the condition on the (3 × n2 )-array tells us more: There is no special role for the elements. We may also permute the lines of the array (as a whole) and still preserve the conditions on the line array and hence obtain a Latin square. Latin squares that are connected by any such permutation are called conjugates. Here is the observation which will make the proof transparent: A partial Latin square obviously corresponds to a partial line array (every pair appears at most once in any two lines), and any conjugate of a partial Latin square is again a partial Latin square. In particular, a partial Latin square can be completed if and only if any conjugate can be completed (just complete the conjugate and then reverse the permutation of the three lines). We will need two results, due to Herbert J. Ryser and to Charles C. Lindner, that were known prior to Smetaniuk’s theorem. If a partial Latin square is of the form that the first r rows are completely filled and the remaining cells are empty, then we speak of an (r × n)-Latin rectangle. Lemma 1. Any (r × n)-Latin rectangle, r < n, can be extended to an ((r+1)×n)-Latin rectangle and hence can be completed to a Latin square.  Proof. We apply Hall’s theorem (see Chapter 27). Let Aj be the set of numbers that do not appear in column j. An admissible (r + 1)-st row corresponds then precisely to a system of distinct representatives for the collection A1 , . . . , An . To prove the lemma we therefore have to verify Hall’s condition (H). Every set Aj has size n − r, and every element is in precisely n − r sets Aj (since it appears r times in the rectangle). Any m of the sets Aj contain together m(n − r) elements and therefore at least m different ones, which is just condition (H).  Lemma 2. Let P be a partial Latin square of order n with at most n − 1 cells filled and at most n2 distinct elements, then P can be completed to a Latin square of order n.

Completing Latin squares

215

 Proof. We first transform the problem into a more convenient form. By the conjugacy principle discussed above, we may replace the condition “at most n2 distinct elements” by the condition that the entries appear in at most n2 rows, and we may further assume that these rows are the top rows. So let the rows with filled

cells be the rows 1, 2, . . . , r, with fi filled r cells in row i, where r ≤ n2 and i=1 fi ≤ n − 1. By permuting the rows, we may assume that f1 ≥ f2 ≥ . . . ≥ fr . Now we complete the rows 1, . . . , r step by step until we reach an (r × n)-rectangle which can then be extended to a Latin square by Lemma 1. Suppose we have already filled rows 1, 2, . . . ,  − 1. In row  there are f filled cells which we may assume to be at the end. The current situation is depicted in the figure, where the shaded part indicates the filled cells. The completion of row  is performed by another application of Hall’s theorem, but this time it is quite subtle. Let X be the set of elements that do not appear in row , thus |X| = n − f , and for j = 1, . . . , n − f let Aj denote the set of those elements in X which do not appear in column j (neither above nor below row ). Hence in order to complete row  we must verify condition (H) for the collection A1 , . . . , An−f . First we claim (1) n − f −  + 1 >  − 1 + f+1 + . . . + fr .

The case  = 1 is clear. Otherwise ri=1 fi < n, f1 ≥ . . . ≥ fr and 1 <  ≤ r together imply n >

r 

fi ≥ ( − 1)f−1 + f + . . . + fr .

i=1

Now either f−1 ≥ 2 (in which case (1) holds) or f−1 = 1. In the latter case, (1) reduces to n > 2( − 1) + r −  + 1 = r +  − 1, which is true because of  ≤ r ≤ n2 . Let us now take m sets Aj , 1 ≤ m ≤ n − f , and let B be their union. We must show |B| ≥ m. Consider the number c of cells in the m columns corresponding to the Aj ’s which contain elements of X. There are at most ( − 1)m such cells above row  and at most f+1 + . . . + fr below row , and thus c ≤ ( − 1)m + f+1 + . . . + fr . On the other hand, each element x ∈ X\B appears in each of the m columns, hence c ≥ m(|X| − |B|), and therefore (with |X| = n − f ) |B| ≥ |X| −

1 mc

≥ n − f − ( − 1) −

1 m (f+1

+ . . . + fr ).

It follows that |B| ≥ m if n − f − ( − 1) −

1 m (f+1

+ . . . + fr ) > m − 1,

that is, if m(n − f −  + 2 − m) > f+1 + . . . + fr .

(2)

A situation for n = 8, with  = 3, f1 = f2 = f3 = 2, f4 = 1. The dark squares represent the pre-filled cells, the lighter ones show the cells that have been filled in the completion process.

216

Completing Latin squares Inequality (2) is true for m = 1 and for m = n−f −+1 by (1), and hence for all values m between 1 and n − f −  + 1, since the left-hand side is a quadratic function in m with leading coefficient −1. The remaining case is m > n − f −  + 1. Since any element x of X is contained in at most  − 1 + f+1 + . . . + fr rows, it can also appear in at most that many columns. Invoking (1) once more, we find that x is in one of the sets Aj , so in this case B = X, |B| = n − f ≥ m, and the proof is complete.  Let us finally prove Smetaniuk’s theorem. Theorem. Any partial Latin square of order n with at most n − 1 filled cells can be completed to a Latin square of the same order.

s1

2

s2

7 5

s3

4

5

s4

4

2

7

4

5 5

4

2

3

4

1

6

5

5

6

1

4

2

3

1

2

3

6

5

4

6

4

5

2

3

1

3

1

6

5

4

2

4

5

2

3

1

6

 Proof. We use induction on n, the cases n ≤ 2 being trivial. Thus we now study a partial Latin square of order n ≥ 3 with at most n − 1 filled cells. With the notation used above these cells lie in r ≤ n − 1 different rows

r numbered s1 , . . . , sr , which contain f1 , . . . , fr > 0 filled cells, with i=1 fi ≤ n − 1. By Lemma 2 we may assume that there are more than n different elements; thus there is an element that appears only once: after 2 renumbering and permutation of rows (if necessary) we may assume that the element n occurs only once, and this is in row s1 . In the next step we want to permute the rows and columns of the partial Latin square such that after the permutations all the filled cells lie below the diagonal — except for the cell filled with n, which will end up on the diagonal. (The diagonal consists of the cells (k, k) with 1 ≤ k ≤ n.) We achieve this as follows: First we permute row s1 into the position f1 . By permutation of columns we move all the filled cells to the left, so that n occurs as the last element in its row, on the diagonal. Next we move row s2 into position 1 + f1 + f2 , and again the filled cells as far to the left as possible. In general, for 1 < i ≤ r we move the row si into position 1 + f1 + f2 + . . . + fi and the filled cells as far left as possible. This clearly gives the desired set-up. The drawing to the left shows an example, with n = 7: the rows s1 = 2, s2 = 3, s3 = 5 and s4 = 7 with f1 = f2 = 2 and f3 = f4 = 1 are moved into the rows numbered 2, 5, 6 and 7, and the columns are permuted “to the left” so that in the end all entries except for the single 7 come to lie below the diagonal, which is marked by •s. In order to be able to apply induction we now remove the entry n from the diagonal and ignore the first row and the last column (which do not not contain any filled cells): thus we are looking at a partial Latin square of order n − 1 with at most n − 2 filled cells, which by induction can be completed to a Latin square of order n − 1. The margin shows one (of many) completions of the partial Latin square that arises in our example. In the figure, the original entries are printed bold. They are already final, as are all the elements in shaded cells; some of the other entries will be changed in the following, in order to complete the Latin square of order n. In the next step we want to move the diagonal elements of the square to the last column and put entries n onto the diagonal in their place. However, in general we cannot do this, since the diagonal elements need not

Completing Latin squares

217

be distinct. Thus we proceed more carefully and perform successively, for k = 2, 3, . . . , n − 1 (in this order), the following operation: Put the value n into the cell (k, n). This yields a correct partial Latin square. Now exchange the value xk in the diagonal cell (k, k) with the value n in the cell (k, n) in the last column. If the value xk did not already occur in the last column, then our job for the current k is completed. After this, the current elements in the k-th column will not be changed any more. In our example this works without problems for k = 2, 3 and 4, and the corresponding diagonal elements 3, 1 and 6 move to the last column. The following three figures show the corresponding operations.

2

7

4

1

6

5

3

2

7

4

1

6

5

3

3

5

6

1

4

2

3

7

5

6

7

4

2

3

1

5

4

1

2

3

6

5

4

1

2

3

6

5

4

7

2

3

1

6

4

5

2

3

1

6

4

5

2

3

1

6

5

4

2

3

1

6

5

4

2

3

1

6

5

4

2

2

3

1

6

4

5

2

3

1

6

4

5

2

3

1

6

2

3

4

1

6

5

5

6

1

4

2

1

2

3

6

6

4

5

3

1

4

5

7

Now we have to treat the case in which there is already an element xk in the last column. In this case we proceed as follows: If there is already an element xk in a cell (j, n) with 2 ≤ j < k, then we exchange in row j the element xk in the n-th column with the element x k in the k-th column. If the element x k also occurs in a cell (j , n), then we also exchange the elements in the j -th row that occur in the n-th and in the k-th columns, and so on. If we proceed like this there will never be two equal entries in a row. Our exchange process ensures that there also will never be two equal elements in a column. So we only have to verify that the exchange process between the k-th and the n-th column does not lead to an infinite loop. This can be seen from the following bipartite graph Gk : Its vertices correspond to the cells (i, k) and (j, n) with 2 ≤ i, j ≤ k whose elements might be exchanged. There is an edge between (i, k) and (j, n) if these two cells lie in the same row (that is, for i = j), or if the cells before the exchange process contain the same element (which implies i = j). In our sketch the edges for i = j are dotted, the others are not. All vertices in Gk have degree 1 or 2. The cell (k, n) corresponds to a vertex of degree 1; this vertex is the beginning of a path which leads to column k on a horizontal edge, then possibly on a sloped edge back to column n, then horizontally back to column k and so on. It ends in column k at a value that does not occur in column n. Thus the exchange operations will end at some point with a step where we move a new element into the last column. Then the work on column k is completed, and the elements in the cells (i, k) for i ≥ 2 are fixed for good.

Gk :

k

n

x k

xk j

x k

x k j

xk

n

k

218

Completing Latin squares In our example the “exchange case” happens for k = 5: the element x5 = 3 does already occur in the last column, so that entry has to be moved back to column k = 5. But the exchange element x 5 = 6 is not new either, it is exchanged by x 5 = 5, and this one is new.

2

7

4

1

6

5

3

2

7

4

1

3

5

6

5

6

7

4

2

3

1

5

6

7

4

2

3

1

1

2

3

7

5

4

6

1

2

3

7

6

4

5

6

4

5

2

3

1

7

6

4

5

2

7

1

3

3

1

6

5

4

2

3

1

6

5

4

2

4

5

2

3

1

6

4

5

2

3

1

6

Finally, the exchange for k = 6 = n − 1 poses no problem, and after that the completion of the Latin square is unique: 7

3

1

6

4

2

4

2

7

4

1

3

5

6

2

7

4

1

3

5

6

2

7

4

1

3

5

6

5

6

7

4

2

3

1

5

6

7

4

2

3

1

5

6

7

4

2

3

1

1

2

3

7

6

4

5

1

2

3

7

6

4

5

1

2

3

7

6

4

5

6

4

5

2

7

1

3

6

4

5

2

7

1

3

6

4

5

2

7

1

3

3

1

6

5

4

2

7

3

1

6

5

4

7

2

3

1

6

5

4

7

2

4

5

2

3

1

6

4

5

2

3

1

6

4

5

2

3

1

6

7

. . . and the same occurs in general: We put an element n into the cell (n, n), and after that the first row can be completed by the missing elements of the respective columns (see Lemma 1), and this completes the proof. In order to get explicitly a completion of the original partial Latin square of order n, we only have to reverse the element, row and column permutations of the first two steps of the proof. 

References [1] T. E VANS : Embedding incomplete Latin squares, Amer. Math. Monthly 67 (1960), 958-961. [2] C. C. L INDNER : On completing Latin rectangles, Canadian Math. Bulletin 13 (1970), 65-68. [3] H. J. RYSER : A combinatorial theorem with an application to Latin rectangles, Proc. Amer. Math. Soc. 2 (1951), 550-552. [4] B. S METANIUK : A new construction on Latin squares I: A proof of the Evans conjecture, Ars Combinatoria 11 (1981), 155-172.

Graph Theory

33 The Dinitz problem 221 34 Five-coloring plane graphs 227 35 How to guard a museum 231 36 Turán’s graph theorem 235 37 Communicating without errors 241 38 The chromatic number of Kneser graphs 251 39 Of friends and politicians 257 40 Probability makes counting (sometimes) easy 261

“The four-colorist geographer”

The Dinitz problem

Chapter 33

The four-color problem was a main driving force for the development of graph theory as we know it today, and coloring is still a topic that many graph theorists like best. Here is a simple-sounding coloring problem, raised by Jeff Dinitz in 1978, which defied all attacks until its astonishingly simple solution by Fred Galvin fifteen years later.

j ↓

Consider n2 cells arranged in an (n × n)-square, and let (i, j) denote the cell in row i and column j. Suppose that for every cell (i, j) we are given a set C(i, j) of n colors. Is it then always possible to color the whole array by picking for each cell (i, j) a color from its set C(i, j) such that the colors in each row and each column are distinct? As a start consider the case when all color sets C(i, j) are the same, say {1, 2, . . . , n}. Then the Dinitz problem reduces to the following task: Fill the (n × n)-square with the numbers 1, 2, . . . , n in such a way that the numbers in any row and column are distinct. In other words, any such coloring corresponds to a Latin square, as discussed in the previous chapter. So, in this case, the answer to our question is “yes.” Since this is so easy, why should it be so much harder in the general case when the set C := i,j C(i, j) contains even more than n colors? The difficulty derives from the fact that not every color of C is available at each cell. For example, whereas in the Latin square case we can clearly choose an arbitrary permutation of the colors for the first row, this is not so anymore in the general problem. Already the case n = 2 illustrates this difficulty. Suppose we are given the color sets that are indicated in the figure. If we choose the colors 1 and 2 for the first row, then we are in trouble since we would then have to pick color 3 for both cells in the second row. Before we tackle the Dinitz problem, let us rephrase the situation in the language of graph theory. As usual we only consider graphs G = (V, E) without loops and multiple edges. Let χ(G) denote the chromatic number of the graph, that is, the smallest number of colors that one can assign to the vertices such that adjacent vertices receive different colors. In other words, a coloring calls for a partition of V into classes (colored with the same color) such that there are no edges within a class. Calling a set A ⊆ V independent if there are no edges within A, we infer that the chromatic number is the smallest number of independent sets which partition the vertex set V. M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_33, © Springer-Verlag Berlin Heidelberg 2013

C(i, j)

← i

{1, 2} {2, 3} {1, 3} {2, 3}

222

The Dinitz problem In 1976 Vizing, and three years later Erd˝os, Rubin, and Taylor, studied the following coloring variant which leads us straight to the Dinitz problem. Suppose in the graph G = (V, E) we are given a set C(v) of colors for each vertex v. A list coloring is a coloring c : V −→ v∈V C(v) where c(v) ∈ C(v) for each v ∈ V . The definition of the list chromatic number χ (G) should now be clear: It is the smallest number k such for any list of color sets C(v) with |C(v)| = k for all v ∈ V there always exists a list coloring. Of course, we have χ (G) ≤ |V | (we never run out of colors). Since ordinary coloring is just the special case of list coloring when all sets C(v) are equal, we obtain for any graph G χ(G) ≤ χ (G). To get back to the Dinitz problem, consider the graph Sn which has as vertex set the n2 cells of our (n × n)-array with two cells adjacent if and only if they are in the same row or column. Since any n cells in a row are pairwise adjacent we need at least n colors. Furthermore, any coloring with n colors corresponds to a Latin square, with the cells occupied by the same number forming a color class. Since Latin squares, as we have seen, exist, we infer χ(Sn ) = n, and the Dinitz problem can now be succinctly stated as

The graph S3

χ (Sn ) = n? {1, 3} {1, 2} {1, 4}

{3, 4}

{2, 3} {2, 4}

One might think that perhaps χ(G) = χ (G) holds for any graph G, but this is a long shot from the truth. Consider the graph G = K2,4 . The chromatic number is 2 since we may use one color for the two left vertices and the second color for the vertices on the right. But now suppose that we are given the color sets indicated in the figure. To color the left vertices we have the four possibilities 1|3, 1|4, 2|3 and 2|4, but any one of these pairs appears as a color set on the right-hand side, so a list coloring is not possible. Hence χ (G) ≥ 3, and the reader may find it fun to prove χ (G) = 3 (there is no need to try out all possibilities!). Generalizing this example, it is not hard to find graphs G where χ(G) = 2, but χ (G) is arbitrarily large! So the list coloring problem is not as easy as it looks at first glance. Back to the Dinitz problem. A significant step towards the solution was made by Jeanette Janssen in 1992 when she proved χ (Sn ) ≤ n + 1, and the coup de grâce was delivered by Fred Galvin by ingeniously combining two results, both of which had long been known. We are going to discuss these two results and show then how they imply χ (Sn ) = n. First we fix some notation. Suppose v is a vertex of the graph G, then we denote as before by d(v) the degree of v. In our square graph Sn every vertex has degree 2n − 2, accounting for the n − 1 other vertices in the same row and in the same column. For a subset A ⊆ V we denote by GA the subgraph which has A as vertex set and which contains all edges of G between vertices of A. We call GA the subgraph induced by A, and say that H is an induced subgraph of G if H = GA for some A.

The Dinitz problem

223

 = (V, E), that is, graphs To state our first result we need directed graphs G where every edge e has an orientation. The notation e = (u, v) means that there is an arc e, also denoted by u−→v, whose initial vertex is u and whose terminal vertex is v. It then makes sense to speak of the outdegree d+ (v) resp. the indegree d− (v), where d+ (v) counts the number of edges with v as initial vertex, and similarly for d− (v); furthermore, d+ (v)+d− (v) = d(v).  without the orientations. When we write G, we mean the graph G The following concept originated in the analysis of games and will play a crucial role in our discussion.  = (V, E) be a directed graph. A kernel K ⊆ V is a Definition 1. Let G subset of the vertices such that (i) K is independent in G, and (ii) for every u ∈ K there exists a vertex v ∈ K with an edge u −→ v. Let us look at the example in the figure. The set {b, c, f } constitutes a kernel, but the subgraph induced by {a, c, e} does not have a kernel since the three edges cycle through the vertices. With all these preparations we are ready to state the first result.  = (V, E) be a directed graph, and suppose that for each Lemma 1. Let G vertex v ∈ V we have a color set C(v) that is larger than the outdegree,  possesses a kernel, |C(v)| ≥ d+ (v) + 1. If every induced subgraph of G then there exists a list coloring of G with a color from C(v) for each v.  Proof. We proceed by induction on |V |. For |V | = 1 there is nothing to prove. Choose a color c ∈ C = v∈V C(v) and set A(c) := {v ∈ V : c ∈ C(v)}. By hypothesis, the induced subgraph GA(c) possesses a kernel K(c). Now we color all v ∈ K(c) with the color c (this is possible since K(c) is independent), and delete K(c) from G and c from C. Let G be the induced subgraph of G on V \K(c) with C (v) = C(v)\c as the new list of color sets. Notice that for each v ∈ A(c)\K(c), the outdegree d+ (v) is decreased by at least 1 (due to condition (ii) of a kernel). So d+ (v) + 1 ≤ |C (v)| still  . The same condition also holds for the vertices outside A(c), holds in G since in this case the color sets C(v) remain unchanged. The new graph G contains fewer vertices than G, and we are done by induction.  The method of attack for the Dinitz problem is now obvious: We have to find an orientation of the graph Sn with outdegrees d+ (v) ≤ n − 1 for all v and which ensures the existence of a kernel for all induced subgraphs. This is accomplished by our second result. Again we need a few preparations. Recall (from Chapter 10) that a bipartite graph G = (X ∪ Y, E) is a graph with the following property: The vertex set V is split into two parts X and Y such that every edge has one endvertex in X and the other in Y . In other words, the bipartite graphs are precisely those which can be colored with two colors (one for X and one for Y ).

a

b

f

c

e

d

224

X

The Dinitz problem

Y

A bipartite graph with a matching

Now we come to an important concept, “stable matchings,” with a downto-earth interpretation. A matching M in a bipartite graph G = (X ∪ Y, E) is a set of edges such that no two edges in M have a common endvertex. In the displayed graph the edges drawn in bold lines constitute a matching. Consider X to be a set of men and Y a set of women and interpret uv ∈ E to mean that u and v might marry. A matching is then a mass-wedding with no person committing bigamy. For our purposes we need a more refined (and more realistic?) version of a matching, suggested by David Gale and Lloyd S. Shapley. Clearly, in real life every person has preferences, and this is what we add to the set-up. In G = (X ∪ Y, E) we assume that for every v ∈ X ∪ Y there is a ranking of the set N (v) of vertices adjacent to v, N (v) = {z1 > z2 > · · · > zd(v) }. Thus z1 is the top choice for v, followed by z2 , and so on. Definition 2. A matching M of G = (X ∪ Y, E) is called stable if the following condition holds: Whenever uv ∈ E\M , u ∈ X, v ∈ Y , then either uy ∈ M with y > v in N (u) or xv ∈ M with x > u in N (v), or both. In our real life interpretation a set of marriages is stable if it never happens that u and v are not married but u prefers v to his partner (if he has one at all) and v prefers u to her mate (if she has one at all), which would clearly be an unstable situation. Before proving our second result let us take a look at the following example:

The bold edges constitute a stable matching. In each priority list, the choice leading to a stable matching is printed bold.

{A > C}

a

A

{c > d > a}

{C > D > B}

b

B

{b}

{A > D}

c

C

{a > b}

{A}

d

D

{c > b}

Notice that in this example there is a unique largest matching M with four edges, M = {aC, bB, cD, dA}, but M is not stable (consider cA). Lemma 2. A stable matching always exists.  Proof. Consider the following algorithm. In the first stage all men u ∈ X propose to their top choice. If a girl receives more than one proposal she picks the one she likes best and keeps him on a string, and if she receives just one proposal she keeps that one on a string. The remaining men are rejected and form the reservoir R. In the second stage all men in R propose to their next choice. The women compare the proposals (together with the one on the string, if there is one), pick their favorite and put him on the string. The rest is rejected and forms the new set R. Now the men in R propose to their next choice, and so on. A man who has proposed to his last choice and is again rejected drops out from further consideration (as well as from the reservoir). Clearly, after some time the reservoir R is empty, and at this point the algorithm stops.

The Dinitz problem

225

Claim. When the algorithm stops, then the men on the strings together with the corresponding girls form a stable matching. Notice first that the men on the string of a particular girl move there in increasing preference (of the girl) since at each stage the girl compares the new proposals with the present mate and then picks the new favorite. Hence if uv ∈ E but uv ∈ M , then either u never proposed to v in which case he found a better mate before he even got around to v, implying uy ∈ M with y > v in N (u), or u proposed to v but was rejected, implying xv ∈ M with x > u in N (v). But this is exactly the condition of a stable matching.  Putting Lemmas 1 and 2 together, we now get Galvin’s solution of the Dinitz problem. Theorem. We have χ (Sn ) = n for all n.  Proof. As before we denote the vertices of Sn by (i, j), 1 ≤ i, j ≤ n. Thus (i, j) and (r, s) are adjacent if and only if i = r or j = s. Take any Latin square L with letters from {1, 2, . . . , n} and denote by L(i, j) n by orienting the entry in cell (i, j). Next make Sn into a directed graph S the horizontal edges (i, j) −→ (i, j ) if L(i, j) < L(i, j ) and the vertical edges (i, j) −→ (i , j) if L(i, j) > L(i , j). Thus, horizontally we orient from the smaller to the larger element, and vertically the other way around. (In the margin we have an example for n = 3.) Notice that we obtain d+ (i, j) = n − 1 for all (i, j). In fact, if L(i, j) = k, then n − k cells in row i contain an entry larger than k, and k − 1 cells in column j have an entry smaller than k. n posBy Lemma 1 it remains to show that every induced subgraph of S sesses a kernel. Consider a subset A ⊆ V , and let X be the set of rows of L, and Y the set of its columns. Associate to A the bipartite graph G = (X ∪ Y, A), where every (i, j) ∈ A is represented by the edge ij with i ∈ X, j ∈ Y . In the example in the margin the cells of A are shaded. The orientation on Sn naturally induces a ranking on the neighborhoods in n respecG = (X ∪ Y, A) by setting j > j in N (i) if (i, j) −→ (i, j ) in S tively i > i in N (j) if (i, j) −→ (i , j). By Lemma 2, G = (X ∪ Y, A) possesses a stable matching M . This M , viewed as a subset of A, is our desired kernel! To see why, note first that M is independent in A since as edges in G = (X ∪ Y, A) they do not share an endvertex i or j. Secondly, if (i, j) ∈ A\M , then by the definition of a stable matching there either n exists (i, j ) ∈ M with j > j or (i , j) ∈ M with i > i, which for S means (i, j) −→ (i, j ) ∈ M or (i, j) −→ (i , j) ∈ M , and the proof is complete.  To end the story let us go a little beyond. The reader may have noticed that the graph Sn arises from a bipartite graph by a simple construction. Take the complete bipartite graph, denoted by Kn,n , with |X| = |Y | = n, and all edges between X and Y . If we consider the edges of Kn,n as vertices

1

2

3

3

1

2

2

3

1

1

2

3

4 1

1

2

2

3

3

4

4

226

The Dinitz problem

G:

L(G) : a b

a

b

d

d c

c

Construction of a line graph

of a new graph, joining two such vertices if and only if as edges in Kn,n they have a common endvertex, then we clearly obtain the square graph Sn . Let us say that Sn is the line graph of Kn,n . Now this same construction can be performed on any graph G with the resulting graph called the line graph L(G) of G. In general, call H a line graph if H = L(G) for some graph G. Of course, not every graph is a line graph, an example being the graph K2,4 that we considered earlier, and for this graph we have seen χ(K2,4 ) < χ (K2,4 ). But what if H is a line graph? By adapting the proof of our theorem it can easily be shown that χ(H) = χ (H) holds whenever H is the line graph of a bipartite graph, and the method may well go some way in verifying the supreme conjecture in this field: Does χ(H) = χ (H) hold for every line graph H? Very little is known about this conjecture, and things look hard — but after all, so did the Dinitz problem twenty years ago.

References ˝ , A. L. RUBIN & H. TAYLOR : Choosability in graphs, Proc. West [1] P. E RD OS Coast Conference on Combinatorics, Graph Theory and Computing, Congressus Numerantium 26 (1979), 125-157.

[2] D. G ALE & L. S. S HAPLEY: College admissions and the stability of marriage, Amer. Math. Monthly 69 (1962), 9-15. [3] F. G ALVIN : The list chromatic index of a bipartite multigraph, J. Combinatorial Theory, Ser. B 63 (1995), 153-158. [4] J. C. M. JANSSEN : The Dinitz problem solved for rectangles, Bulletin Amer. Math. Soc. 29 (1993), 243-249. [5] V. G. V IZING : Coloring the vertices of a graph in prescribed colours (in Russian), Metody Diskret. Analiz. 101 (1976), 3-10.

Five-coloring plane graphs

Plane graphs and their colorings have been the subject of intensive research since the beginnings of graph theory because of their connection to the fourcolor problem. As stated originally the four-color problem asked whether it is always possible to color the regions of a plane map with four colors such that regions which share a common boundary (and not just a point) receive different colors. The figure on the right shows that coloring the regions of a map is really the same task as coloring the vertices of a plane graph. As in Chapter 12 (page 75) place a vertex in the interior of each region (including the outer region) and connect two such vertices belonging to neighboring regions by an edge through the common boundary. The resulting graph G, the dual graph of the map M , is then a plane graph, and coloring the vertices of G in the usual sense is the same as coloring the regions of M . So we may as well concentrate on vertex-coloring plane graphs and will do so from now on. Note that we may assume that G has no loops or multiple edges, since these are irrelevant for coloring.

Chapter 34

The dual graph of a map

In the long and arduous history of attacks to prove the four-color theorem many attempts came close, but what finally succeeded in the Appel–Haken proof of 1976 and also in the more recent proof of Robertson, Sanders, Seymour and Thomas 1997 was a combination of very old ideas (dating back to the 19th century) and the very new calculating powers of modernday computers. Twenty-five years after the original proof, the situation is still basically the same, there is even a computer-generated computercheckable proof due to Gonthier, but no proof from The Book is in sight. So let us be more modest and ask whether there is a neat proof that every plane graph can be 5-colored. A proof of this five-color theorem had already been given by Heawood at the turn of the century. The basic tool for his proof (and indeed also for the four-color theorem) was Euler’s formula (see Chapter 12). Clearly, when coloring a graph G we may assume that G is connected since we may color the connected pieces separately. A plane graph divides the plane into a set R of regions (including the exterior region). Euler’s formula states that for plane connected graphs G = (V, E) we always have |V | − |E| + |R| = 2. As a warm-up, let us see how Euler’s formula may be applied to prove that every plane graph G is 6-colorable. We proceed by induction on the number n of vertices. For small values of n (in particular, for n ≤ 6) this is obvious. M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_34, © Springer-Verlag Berlin Heidelberg 2013

This plane graph has 8 vertices, 13 edges and 7 regions.

228

Five-coloring plane graphs From part (A) of the proposition on page 77 we know that G has a vertex v of degree at most 5. Delete v and all edges incident with v. The resulting graph G = G\v is a plane graph on n − 1 vertices. By induction, it can be 6-colored. Since v has at most 5 neighbors in G, at most 5 colors are used for these neighbors in the coloring of G . So we can extend any 6-coloring of G to a 6-coloring of G by assigning a color to v which is not used for any of its neighbors in the coloring of G . Thus G is indeed 6-colorable. Now let us look at the list chromatic number of plane graphs, which we have discussed in the chapter on the Dinitz problem. Clearly, our 6-coloring method works for lists of colors as well (again we never run out of colors), so χ (G) ≤ 6 holds for any plane graph G. Erd˝os, Rubin and Taylor conjectured in 1979 that every plane graph has list chromatic number at most 5, and further that there are plane graphs G with χ (G) > 4. They were right on both counts. Margit Voigt was the first to construct an example of a plane graph G with χ (G) = 5 (her example had 238 vertices) and around the same time Carsten Thomassen gave a truly stunning proof of the 5-list coloring conjecture. His proof is a telling example of what you can do when you find the right induction hypothesis. It does not use Euler’s formula at all!

Theorem. All planar graphs G can be 5-list colored: χ (G) ≤ 5.

 Proof. First note that adding edges can only increase the chromatic number. In other words, when H is a subgraph of G, then χ (H) ≤ χ (G) certainly holds. Hence we may assume that G is connected and that all the bounded faces of an embedding have triangles as boundaries. Let us call such a graph near-triangulated. The validity of the theorem for neartriangulated graphs will establish the statement for all plane graphs. The trick of the proof is to show the following stronger statement (which allows us to use induction): Let G = (V, E) be a near-triangulated graph, and let B be the cycle bounding the outer region. We make the following assumptions on the color sets C(v), v ∈ V : A near-triangulated plane graph

(1) Two adjacent vertices x, y of B are already colored with (different) colors α and β. (2) |C(v)| ≥ 3 for all other vertices v of B. (3) |C(v)| ≥ 5 for all vertices v in the interior. Then the coloring of x, y can be extended to a proper coloring of G by choosing colors from the lists. In particular, χ (G) ≤ 5.

Five-coloring plane graphs For |V | = 3 this is obvious, since for the only uncolored vertex v we have |C(v)| ≥ 3, so there is a color available. Now we proceed by induction. Case 1: Suppose B has a chord, that is, an edge not in B that joins two vertices u, v ∈ B. The subgraph G1 which is bounded by B1 ∪ {uv} and contains x, y, u and v is near-triangulated and therefore has a 5-list coloring by induction. Suppose in this coloring the vertices u and v receive the colors γ and δ. Now we look at the bottom part G2 bounded by B2 and uv. Regarding u, v as pre-colored, we see that the induction hypotheses are also satisfied for G2 . Hence G2 can be 5-list colored with the available colors, and thus the same is true for G. Case 2: Suppose B has no chord. Let v0 be the vertex on the other side of the α-colored vertex x on B, and let x, v1 , . . . , vt , w be the neighbors of v0 . Since G is near-triangulated we have the situation shown in the figure. Construct the near-triangulated graph G = G\v0 by deleting from G the vertex v0 and all edges emanating from v0 . This G has as outer boundary B = (B\v0 ) ∪ {v1 , . . . , vt }. Since |C(v0 )| ≥ 3 by assumption (2) there exist two colors γ, δ in C(v0 ) different from α. Now we replace every color set C(vi ) by C(vi )\{γ, δ}, keeping the original color sets for all other vertices in G . Then G clearly satisfies all assumptions and is thus 5-list colorable by induction. Choosing γ or δ for v0 , different from the color of w, we can extend the list coloring of G to all of G. 

229 B1

x y

G1 u

v G2 B2

v0 w

x(α) vt . . .

y(β) v2

v1

B

So, the 5-list color theorem is proved, but the story is not quite over. A stronger conjecture claimed that the list-chromatic number of a plane graph G is at most 1 more than the ordinary chromatic number: Is χ (G) ≤ χ(G) + 1 for every plane graph G ? Since χ(G) ≤ 4 by the four-color theorem, we have three cases: Case I: χ(G) = 2 =⇒ χ (G) ≤ 3 Case II: χ(G) = 3 =⇒ χ (G) ≤ 4

α

Case III: χ(G) = 4 =⇒ χ (G) ≤ 5. Thomassen’s result settles Case III, and Case I was proved by an ingenious (and much more sophisticated) argument by Alon and Tarsi. Furthermore, there are plane graphs G with χ(G) = 2 and χ (G) = 3, for example the graph K2,4 that we considered in the chapter on the Dinitz problem. But what about Case II? Here the conjecture fails: This was first shown by Margit Voigt for a graph that was earlier constructed by Shai Gutner. His graph on 130 vertices can be obtained as follows. First we look at the “double octahedron” (see the figure), which is clearly 3-colorable. Let α ∈ {5, 6, 7, 8} and β ∈ {9, 10, 11, 12}, and consider the lists that are given in the figure. You are invited to check that with these lists a coloring is not possible. Now take 16 copies of this graph, and identify all top vertices and all bottom vertices. This yields a graph on 16 · 8 + 2 = 130 vertices which

{α, 1, 3, 4}

{α, β, 1, 2}

{α, 2, 3, 4}

{1, 2, 3, 4} {1, 2, 3, 4}

{β, 2, 3, 4}

{α, β, 1, 2}

{β, 1, 3, 4} β

230

Five-coloring plane graphs is still plane and 3-colorable. We assign {5, 6, 7, 8} to the top vertex and {9, 10, 11, 12} to the bottom vertex, with the inner lists corresponding to the 16 pairs (α, β), α ∈ {5, 6, 7, 8}, β ∈ {9, 10, 11, 12}. For every choice of α and β we thus obtain a subgraph as in the figure, and so a list coloring of the big graph is not possible. By modifying another one of Gutner’s examples, Voigt and Wirth came up with an even smaller plane graph with 75 vertices and χ = 3, χ = 5, which in addition uses only the minimal number of 5 colors in the combined lists. The current record is 63 vertices.

References [1] N. A LON & M. TARSI : Colorings and orientations of graphs, Combinatorica 12 (1992), 125-134. ˝ , A. L. RUBIN & H. TAYLOR : Choosability in graphs, Proc. West [2] P. E RD OS Coast Conference on Combinatorics, Graph Theory and Computing, Congressus Numerantium 26 (1979), 125-157.

[3] G. G ONTHIER : Formal proof — the Four-Color Theorem, Notices of the AMS (11) 55 (2008), 1382-1393. [4] S. G UTNER : The complexity of planar graph choosability, Discrete Math. 159 (1996), 119-130. [5] N. ROBERTSON , D. P. S ANDERS , P. S EYMOUR & R. T HOMAS : The fourcolour theorem, J. Combinatorial Theory, Ser. B 70 (1997), 2-44. [6] C. T HOMASSEN : Every planar graph is 5-choosable, J. Combinatorial Theory, Ser. B 62 (1994), 180-181. [7] M. VOIGT: List colorings of planar graphs, Discrete Math. 120 (1993), 215-219. [8] M. VOIGT & B. W IRTH : On 3-colorable non-4-choosable planar graphs, J. Graph Theory 24 (1997), 233-235.

How to guard a museum

Here is an appealing problem which was raised by Victor Klee in 1973. Suppose the manager of a museum wants to make sure that at all times every point of the museum is watched by a guard. The guards are stationed at fixed posts, but they are able to turn around. How many guards are needed? We picture the walls of the museum as a polygon consisting of n sides. Of course, if the polygon is convex, then one guard is enough. In fact, the guard may be stationed at any point of the museum. But, in general, the walls of the museum may have the shape of any closed polygon. Consider a comb-shaped museum with n = 3m walls, as depicted on the right. It is easy to see that this requires at least m = n3 guards. In fact, there are n walls. Now notice that the point 1 can only be observed by a guard stationed in the shaded triangle containing 1, and similarly for the other points 2, 3, . . . , m. Since all these triangles are disjoint we conclude that at least m guards are needed. But m guards are also enough, since they can be placed at the top lines of the triangles. By cutting off one or two walls at the end, we conclude that for any n there is an n-walled museum which requires n3 guards.

Chapter 35

A convex exhibition hall

... 1

2

3

A real life art gallery. . . M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_35, © Springer-Verlag Berlin Heidelberg 2013

...

m

232

How to guard a museum The following result states that this is the worst case. Theorem. For any museum with n walls, n3 guards suffice. This “art gallery theorem” was first proved by Vašek Chvátal by a clever argument, but here is a proof due to Steve Fisk that is truly beautiful.

A museum with n = 12 walls

 Proof. First of all, let us draw n − 3 non-crossing diagonals between corners of the walls until the interior is triangulated. For example, we can draw 9 diagonals in the museum depicted in the margin to produce a triangulation. It does not matter which triangulation we choose, any one will do. Now think of the new figure as a plane graph with the corners as vertices and the walls and diagonals as edges. Claim. This graph is 3-colorable.

A triangulation of the museum

C B



A

C A

B

Schönhardt’s polyhedron: The interior dihedral angles at the edges AB  , BC  and CA are greater than 180◦ .

For n = 3 there is nothing to prove. Now for n > 3 pick any two vertices u and v which are connected by a diagonal. This diagonal will split the graph into two smaller triangulated graphs both containing the edge uv. By induction we may color each part with 3 colors where we may choose color 1 for u and color 2 for v in each coloring. Pasting the colorings together yields a 3-coloring of the whole graph. The rest is easy. Since there are n vertices, at least one of the color classes, say the vertices colored 1, contains at most n3 vertices, and this is where we place the guards. Since every triangle contains a vertex of color 1 we infer that every triangle is guarded, and hence so is the whole museum.  The astute reader may have noticed a subtle point in our reasoning. Does a triangulation always exist? Probably everybody’s first reaction is: Obviously, yes! Well, it does exist, but this is not completely obvious, and, in fact, the natural generalization to three dimensions (partitioning into tetrahedra) is false! This may be seen from Schönhardt’s polyhedron, depicted on the left. It is obtained from a triangular prism by rotating the top triangle, so that each of the quadrilateral faces breaks into two triangles with a non-convex edge. Try to triangulate this polyhedron! You will notice that any tetrahedron that contains the bottom triangle must contain one of the three top vertices: but the resulting tetrahedron will not be contained in Schönhardt’s polyhedron. So there is no triangulation without an additional vertex. To prove that a triangulation exists in the case of a planar non-convex polygon, we proceed by induction on the number n of vertices. For n = 3 the polygon is a triangle, and there is nothing to prove. Let n ≥ 4. To use induction, all we have to produce is one diagonal which will split the polygon P into two smaller parts, such that a triangulation of the polygon can be pasted together from triangulations of the parts. Call a vertex A convex if the interior angle at the vertex is less than 180◦ . Since the sum of the interior angles of P is (n − 2)180◦, there must be a

How to guard a museum convex vertex A. In fact, there must be at least three of them: In essence this is an application of the pigeonhole principle! Or you may consider the convex hull of the polygon, and note that all its vertices are convex also for the original polygon. Now look at the two neighboring vertices B and C of A. If the segment BC lies entirely in P , then this is our diagonal. If not, the triangle ABC contains other vertices. Slide BC towards A until it hits the last vertex Z in ABC. Now AZ is within P , and we have a diagonal. There are many variants to the art gallery theorem. For example, we may only want to guard the walls (which is, after all, where the paintings hang), or the guards are all stationed at vertices. A particularly nice (unsolved) variant goes as follows: Suppose each guard may patrol one wall of the museum, so he walks along his wall and sees anything that can be seen from any point along this wall. How many “wall guards” do we then need to keep control? Godfried Toussaint constructed the example of a museum displayed here which shows that n4 guards may be necessary. This polygon has 28 sides (and, in general, 4m sides), and the reader is invited to check that m wall-guards are needed. It is conjectured that, except for some small values of n, this number is also sufficient, but a proof, let alone a Book Proof, is still missing.

References [1] V. C HVÁTAL : A combinatorial theorem in plane geometry, J. Combinatorial Theory, Ser. B 18 (1975), 39-41. [2] S. F ISK : A short proof of Chvátal’s watchman theorem, J. Combinatorial Theory, Ser. B 24 (1978), 374. [3] J. O’ROURKE : Art Gallery Theorems and Algorithms, Oxford University Press 1987. [4] E. S CHÖNHARDT: Über die Zerlegung von Dreieckspolyedern in Tetraeder, Math. Annalen 98 (1928), 309-312.

233

C

B Z A

234

“Museum guards” (A 3-dimensional art-gallery problem)

How to guard a museum

Chapter 36

Turán’s graph theorem

One of the fundamental results in graph theory is the theorem of Turán from 1941, which initiated extremal graph theory. Turán’s theorem was rediscovered many times with various different proofs. We will discuss five of them and let the reader decide which one belongs in The Book. Let us fix some notation. We consider simple graphs G on the vertex set V = {v1 , . . . , vn } and edge set E. If vi and vj are neighbors, then we write vi vj ∈ E. A p-clique in G is a complete subgraph of G on p vertices, denoted by Kp . Paul Turán posed the following question: Suppose G is a simple graph that does not contain a p-clique. What is the largest number of edges that G can have? We readily obtain examples of such graphs by dividing V into p−1 pairwise disjoint subsets V = V1 ∪ . . . ∪ Vp−1 , |Vi | = ni , n = n1 + . . . + np−1 , joining two vertices if and only if they lie in distinct sets Vi , Vj . We denote the resulting graph by Kn1 ,...,np−1 ; it has i w2 > 0, then choose ε with 0 < ε < w1 − w2 and change w1 to w1 − ε and w2 to w2 + ε. The new distribution w satisfies f (w ) = f (w) + ε(w1 − w2 ) − ε2 > f (w), and we infer that the maximal value of f (w) is attained for wi = k1 on a k-clique and wi = 0 otherwise. Since a k-clique contains k(k−1) edges, we obtain 2 1 k(k − 1) 1 1 1 − . f (w) = = 2 k2 2 k Since this expression is increasing in k, the best we can do is to set k = p−1 (since G has no p-cliques). So we conclude 1 1  f (w) ≤ 1− 2 p−1 for any distribution w. In particular, this inequality holds for the uniform distribution given by wi = n1 for all i. Thus we find  |E| 1 1  1 ≤ 1 − , = f w = i n2 n 2 p−1 

which is precisely (1).

 Fourth proof. This time we use some concepts from probability theory. Let G be an arbitrary graph on the vertex set V = {v1 , . . . , vn }. Denote the degree of vi by di , and write ω(G) for the number of vertices in a largest clique, called the clique number of G. Claim. We have ω(G) ≥

n  i=1

1 . n − di

We choose a random permutation π = v1 v2 . . . vn of the vertex set V , where each permutation is supposed to appear with the same probability 1 n! , and then consider the following set Cπ . We put vi into Cπ if and only if vi is adjacent to all vj (j < i) preceding vi . By definition, Cπ is a clique

in G. Let X = |Cπ | be the corresponding random variable. We have n X = i=1 Xi , where Xi is the indicator random variable of the vertex vi , that is, Xi = 1 or Xi = 0 depending on whether vi ∈ Cπ or vi ∈ Cπ . Note that vi belongs to Cπ with respect to the permutation v1 v2 . . . vn if and only if vi appears before all n − 1 − di vertices which are not adjacent to vi , or in other words, if vi is the first among vi and its n − 1 − di non-neighbors. 1 1 The probability that this happens is n−d , hence EXi = n−d . i i

“Moving weights”

238

Turán’s graph theorem Thus by linearity of expectation (see page 94) we obtain E(|Cπ |) = EX =

n 

EXi =

i=1

n  i=1

1 . n − di

Consequently, there must be a clique of at least that size, and this was our claim. To deduce Turán’s theorem from the claim we use the Cauchy– Schwarz inequality from Chapter 18, n 

2 ai b i

≤

i=1

√ Set ai = n − di , bi =

a2i

i=1 √ 1 , n−di

n 

 b2i .

n=1

then ai bi = 1, and so

n n   n2 ≤ ( (n − di ))( i=1

n 

i=1

 1 ) ≤ ω(G) (n − di ). n − di i=1 n

(2)

At this point apply the hypothesis ω(G) ≤ p − 1 of Turán’s theorem.

we n Using also i=1 di = 2|E| from the chapter on double counting, inequality (2) leads to n2 ≤ (p − 1)(n2 − 2|E|), and this is equivalent to Turán’s inequality.



Now we are ready for the last proof, which may be the most beautiful of them all. Its origin is not clear; we got it from Stephan Brandt, who heard it in Oberwolfach. It may be “folklore” graph theory. It yields in one stroke that the Turán graph is in fact the unique example with a maximal number of edges. It may be noted that both proofs 1 and 2 also imply this stronger result.  Fifth proof. Let G be a graph on n vertices without a p-clique and with a maximal number of edges. w Claim. G does not contain three vertices u, v, w such that vw ∈ E, but uv ∈ E, uw ∈ E. v

u

w v

v u

Suppose otherwise, and consider the following cases. Case 1: d(u) < d(v) or d(u) < d(w). We may suppose that d(u) < d(v). Then we duplicate v, that is, we create a new vertex v which has exactly the same neighbors as v (but vv is not an edge), delete u, and keep the rest unchanged. The new graph G has again no p-clique, and for the number of edges we find |E(G )| = |E(G)| + d(v) − d(u) > |E(G)|, a contradiction.

Turán’s graph theorem Case 2: d(u) ≥ d(v) and d(u) ≥ d(w). Duplicate u twice and delete v and w (as illustrated in the margin). Again, the new graph G has no p-clique, and we compute (the −1 results from the edge vw):

239 w v

u

|E(G )| = |E(G)| + 2d(u) − (d(v) + d(w) − 1) > |E(G)|. So we have a contradiction once more. A moment’s thought shows that the claim we have proved is equivalent to the statement that u ∼ v :⇐⇒ uv ∈ E(G) defines an equivalence relation. Thus G is a complete multipartite graph, G = Kn1 ,...,np−1 , and we are finished. 

References [1] M. A IGNER : Turán’s graph theorem, Amer. Math. Monthly 102 (1995), 808-816. [2] N. A LON & J. S PENCER : The Probabilistic Method, Wiley Interscience 1992. ˝ : On the graph theorem of Turán (in Hungarian), Math. Fiz. Lapok [3] P. E RD OS 21 (1970), 249-251.

[4] T. S. M OTZKIN & E. G. S TRAUS : Maxima for graphs and a new proof of a theorem of Turán, Canad. J. Math. 17 (1965), 533-540. [5] P. T URÁN : On an extremal problem in graph theory, Math. Fiz. Lapok 48 (1941), 436-452.

“Larger weights to move”

u u

Communicating without errors

Chapter 37

In 1956, Claude Shannon, the founder of information theory, posed the following very interesting question: Suppose we want to transmit messages across a channel (where some symbols may be distorted) to a receiver. What is the maximum rate of transmission such that the receiver may recover the original message without errors? Let us see what Shannon meant by “channel” and “rate of transmission.” We are given a set V of symbols, and a message is just a string of symbols from V. We model the channel as a graph G = (V, E), where V is the set of symbols, and E the set of edges between unreliable pairs of symbols, that is, symbols which may be confused during transmission. For example, communicating over a phone in everyday language, we connnect the symbols B and P by an edge since the receiver may not be able to distinguish them. Let us call G the confusion graph. The 5-cycle C5 will play a prominent role in our discussion. In this example, 1 and 2 may be confused, but not 1 and 3, etc. Ideally we would like to use all 5 symbols for transmission, but since we want to communicate error-free we can — if we only send single symbols — use only one letter from each pair that might be confused. Thus for the 5-cycle we can use only two different letters (any two that are not connected by an edge). In the language of information theory, this means that for the 5-cycle we achieve an information rate of log2 2 = 1 (instead of the maximal log2 5 ≈ 2.32). It is clear that in this model, for an arbitrary graph G = (V, E), the best we can do is to transmit symbols from a largest independent set. Thus the information rate, when sending single symbols, is log2 α(G), where α(G) is the independence number of G. Let us see whether we can increase the information rate by using larger strings in place of single symbols. Suppose we want to transmit strings of length 2. The strings u1 u2 and v1 v2 can only be confused if one of the following three cases holds: • u1 = v1 and u2 can be confused with v2 , • u2 = v2 and u1 can be confused with v1 , or • u1 = v1 can be confused and u2 = v2 can be confused. In graph-theoretic terms this amounts to considering the product G1 × G2 of two graphs G1 = (V1 , E1 ) and G2 = (V2 , E2 ). G1 × G2 has the vertex M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_37, © Springer-Verlag Berlin Heidelberg 2013

Claude Shannon

1 5

2

4

3

242

Communicating without errors set V1 × V2 = {(u1 , u2 ) : u1 ∈ V1 , u2 ∈ V2 }, with (u1 , u2 ) = (v1 , v2 ) connected by an edge if and only if ui = vi or ui vi ∈ Ei for i = 1, 2. The confusion graph for strings of length 2 is thus G2 = G × G, the product of the confusion graph G for single symbols with itself. The information rate of strings of length 2 per symbol is then given by  log2 α(G2 ) = log2 α(G2 ). 2 Now, of course, we may use strings of any length n. The n-th confusion graph Gn = G×G×. . .×G has vertex set V n = {(u1 , . . . , un ) : ui ∈ V } with (u1 , . . . , un ) = (v1 , . . . vn ) being connected by an edge if ui = vi or ui vi ∈ E for all i. The rate of information per symbol determined by strings of length n is  log2 α(Gn ) = log2 n α(Gn ). n What can we say about α(Gn )? Here is a first observation. Let U ⊆ V be a largest independent set in G, |U | = α. The αn vertices in Gn of the form (u1 , . . . , un ), ui ∈ U for all i, clearly form an independent set in Gn . Hence

and therefore

α(Gn ) ≥

α(G)n

 n α(Gn ) ≥

α(G),

meaning that we never decrease the information rate by using longer strings instead of single symbols. This, by the way, is a basic idea of coding theory: By encoding symbols into longer strings we can make error-free communication more efficient. Disregarding the logarithm we thus arrive at Shannon’s fundamental definition: The zero-error capacity of a graph G is given by  Θ(G) := sup n α(Gn ), n≥1

and Shannon’s problem was to compute Θ(G), and in particular Θ(C5 ). Let us look at C5 . So far we know α(C5 ) = 2 ≤ Θ(C5 ). Looking at the 5-cycle as depicted earlier, or at the product C5 × C5 as drawn on the left, we see that the set {(1, 1), (2, 3), (3, 5), (4, 2), (5, 4)} is independent in C52 . Thus we have α(C52 ) ≥ 5. Since an independent set can contain only two vertices from any two consecutive rows we see that α(C52 ) = 5. Hence, by using strings √ of length 2 we have increased the lower bound for the capacity to Θ(C5 ) ≥ 5.

The graph C5 × C5

So far we have no upper bounds for the capacity. To obtain such bounds we again follow Shannon’s original ideas. First we need the dual definition of an independent set. We recall that a subset C ⊆ V is a clique if any two vertices of C are joined by an edge. Thus the vertices form trivial

Communicating without errors

243

cliques of size 1, the edges are the cliques of size 2, the triangles are cliques of size 3, and so on. Let C be the set of cliques in G. Consider an arbitrary probability distribution x = (xv : v ∈ V ) on the set of vertices, that

is, xv ≥ 0 and v∈V xv = 1. To every distribution x we associate the “maximal value of a clique”  λ(x) = max xv , C∈C

v∈C

and finally we set



λ(G) = min λ(x) = min max x

x

C∈C

xv .

v∈C

To be precise we should use inf instead of min, but the minimum exists because λ(x) is continuous on the compact set of all distributions. Consider now an independent set U ⊆ V of maximal size α(G) = α. Associated to U we define the distribution xU = (xv : v ∈ V ) by setting xv = α1 if v ∈ U and xv = 0 otherwise. Since any clique contains at most one vertex from U , we infer λ(xU ) = α1 , and thus by the definition of λ(G) λ(G) ≤

1 α(G)

or

α(G) ≤ λ(G)−1 .

What Shannon observed is that λ(G)−1 is, in fact, an upper bound for all  n α(Gn ), and hence also for Θ(G). In order to prove this it suffices to show that for graphs G, H λ(G × H) = λ(G)λ(H)

(1)

holds, since this will imply λ(Gn ) = λ(G)n and hence α(Gn ) ≤  n α(Gn ) ≤

λ(Gn )−1 = λ(G)−n λ(G)−1 .

To prove (1) we make use of the duality theorem of linear programming (see [1]) and get   λ(G) = min max xv = max min yC , (2) x

C∈C

y

v∈C

v∈V

Cv

where the right-hand side runs through all probability distributions y = (yC : C ∈ C) on C. Consider G × H, and let x and x be distributions which achieve the minima, λ(x) = λ(G), λ(x ) = λ(H). In the vertex set of G × H we assign the value z(u,v) = xu x v to the vertex (u, v). Since (u,v) z(u,v) =

u xu v xv = 1, we obtain a distribution. Next we observe that the maximal cliques in G × H are of the form C × D = {(u, v) : u ∈ C, v ∈ D} where C and D are cliques in G and H, respectively. Hence we obtain  λ(G × H) ≤ λ(z) = max z(u,v) C×D

=

max

C×D

(u,v)∈C×D



u∈C

xu



v∈D

x v = λ(G)λ(H)

244

Communicating without errors by the definition of λ(G × H). In the same way the converse inequality λ(G × H) ≥ λ(G)λ(H) is shown by using the dual expression for λ(G) in (2). In summary we can state: Θ(G) ≤ λ(G)−1 , for any graph G. Let us apply our findings to the 5-cycle and, more generally, to the 1 1 m-cycle Cm . By using the uniform distribution ( m ,..., m ) on the 2 vertices, we obtain λ(Cm ) ≤ m , since any clique contains at most two 1 vertices. Similarly, choosing m for the edges and 0 for the vertices, we have 2 2 λ(Cm ) ≥ m by the dual expression in (2). We conclude that λ(Cm ) = m and therefore m Θ(Cm ) ≤ 2 for all m. Now, if m is even, then clearly α(Cm ) = m 2 and thus also m−1 Θ(Cm ) = m 2 . For odd m, however, we have α(Cm ) = 2 . For m = 3, C3 is a clique, and so is every product C3n , implying α(C3 ) = Θ(C3 ) = 1. So, the first interesting case is the 5-cycle, where we know up to now √ 5 5 ≤ Θ(C5 ) ≤ . 2

(3)

Using his linear programming approach (and some other ideas) Shannon was able to compute the capacity of many graphs and, in particular, of all graphs with five or fewer vertices — with the single exception of C5 , where he could not go beyond the bounds in (3). This is where things stood for more than 20 years until László Lovász showed by an astonishingly simple √ argument that indeed Θ(C5 ) = 5. A seemingly very difficult combinatorial problem was provided with an unexpected and elegant solution. Lovász’ main new idea was to represent the vertices v of the graph by real vectors of length 1 such that any two vectors which belong to nonadjacent vertices in G are orthogonal. Let us call such a set of vectors an orthonormal representation of G. Clearly, such a representation always exists: just take the unit vectors (1, 0, . . . , 0)T , (0, 1, 0, . . . , 0)T , . . . , (0, 0, . . . , 1)T of dimension m = |V |. For the graph C5 we may obtain an orthonormal representation in R3 by considering an “umbrella” with five ribs v 1 , . . . , v 5 of unit length. Now open the umbrella (with tip at the origin) to the point where the angles between alternate ribs are 90◦ . Lovász then went on to show that the height h of the umbrella, that is, the distance between 0 and S, provides the bound

⎫ S ⎪ ⎪ ⎪ ⎪ ⎪ ⎬ ⎪ ⎪ ⎪ ⎪ ⎪ ⎭

h

0 The Lovász umbrella

Θ(C5 ) ≤

1 . h2

(4)

A simple calculation yields h2 = √15 ; see the box on the next page. From √ √ this Θ(C5 ) ≤ 5 follows, and therefore Θ(C5 ) = 5.

Communicating without errors

245

Let us see how Lovász proceeded to prove the inequality (4). (His results were, in fact, much more general.) Consider the usual inner product x, y = x1 y1 + . . . + xs ys of two vectors x = (x1 , . . . , xs ), y = (y1 , . . . , ys ) in Rs . Then |x|2 = x, x = x21 + . . . + x2s is the square of the length |x| of x, and the angle γ between x and y is given by cos γ =

x, y . |x||y|

Thus x, y = 0 if and only if x and y are orthogonal.

Pentagons and the golden section Tradition has it that a rectangle was considered aesthetically pleasing if, after cutting off a square of length a, the remaining rectangle had the same shape as the original one. The side lengths a, b of such a a rectangle must satisfy ab = b−a . Setting τ := ab for the ratio, we 1 2 obtain τ = τ −1 or τ − τ − 1 = 0. Solving the quadratic equation

b

a

√

yields the golden section τ = 1+2 5 ≈ 1.6180. Consider now a regular pentagon of side length a, and let d be the length of its diagonals. It was already known to Euclid (Book XIII,8) that ad = τ , and that the intersection point of two diagonals divides the diagonals in the golden section. Here is Euclid’s Book Proof. Since the total angle sum of the pentagon is 3π, the angle at any vertex equals 3π 5 . It follows that ABE = π5 , since ABE is an isosceles triangle. This, in turn, implies AM B = 3π 5 , and we conclude that the triangles ABC and AM B are similar. The quadrilateral CM ED is a rhombus since opposing sides are parallel (look at the angles), and so |M C| = a and thus |AM | = d − a. By the similarity of ABC and AM B we conclude d |AC| |AB| a |M C| = = = = = τ. a |AB| |AM | d−a |M A| There is more to come. For the distance s of a vertex to the center of 2 the pentagon S, the reader is invited to prove the relation s2 = τd+2 (note that BS cuts the diagonal AC at a right angle and halves it). To finish our excursion into geometry, consider now the umbrella with the regular pentagon on top. Since alternate ribs (of length 1) √ form a right angle, the theorem of Pythagoras gives us d = 2, and 2 4 hence s2 = τ +2 = √5+5 . So, with Pythagoras again, we find for the height h = |OS| our promised result √ 1 1+ 5 2 2 = √ . h = 1−s = √ 5+5 5

b−a

A

B

E

M S

C

D

246

Communicating without errors Now we head for an upper bound for the Shannon capacity of any graph G that has an especially “nice” orthonormal representation. For this let T = {v (1) , . . . , v (m) } be an orthonormal representation of G in Rs , where v (i) corresponds to the vertex vi . We assume in addition that all the vectors v (i) 1 have the same angle (= 90◦ ) with the vector u := m (v (1) + . . . + v (m) ), or equivalently that the inner product v (i) , u = σT has the same value σT = 0 for all i. Let us call this value σT the constant of the representation T . For the Lovász umbrella that represents C5 the  condition v (i) , u = σT certainly holds, for u = OS. Now we proceed in the following three steps. (A) Consider a probability distribution x = (x1 , . . . , xm ) on V and set μ(x) := |x1 v (1) + . . . + xm v (m) |2 , and μT (G) := inf μ(x). x

Let U be a largest independent set in G with |U | = α, and define xU = (x1 , . . . , xm ) with xi = α1 if vi ∈ U and xi = 0 otherwise. Since all vectors v (i) have unit length and v (i) , v (j)  = 0 for any two non-adjacent vertices, we infer m m &2 &  1 1 & & μT (G) ≤ μ(xU ) = & xi v (i) & = x2i = α 2 = . α α i=1 i=1

Thus we have μT (G) ≤ α−1 , and therefore α(G) ≤

1 . μT (G)

(B) Next we compute μT (G). We need the Cauchy–Schwarz inequality a, b2 ≤ |a|2 |b|2 for vectors a, b ∈ Rs . Applied to a = x1 v (1) + . . . + xm v (m) and b = u, the inequality yields x1 v (1) + . . . + xm v (m) , u2 ≤ μ(x) |u|2 .

(5)

By our assumption that v (i) , u = σT for all i, we have x1 v (1) + . . . + xm v (m) , u = (x1 + . . . + xm ) σT = σT for any distribution x. Thus, in particular, this has to hold for the uniform 1 1 distribution ( m ,..., m ), which implies |u|2 = σT . Hence (5) reduces to σT2 ≤ μ(x) σT

or

μT (G) ≥ σT .

Communicating without errors

247

1 1 On the other hand, for x = ( m ,..., m ) we obtain 1 μT (G) ≤ μ(x) = | m (v (1) + . . . + v (m) )|2 = |u|2 = σT ,

and so we have proved (6)

μT (G) = σT . In summary, we have established the inequality α(G) ≤

1 σT

(7)

for any orthonormal respresentation T with constant σT . (C) To extend this inequality to Θ(G), we proceed as before. Consider again the product G × H of two graphs. Let G and H have orthonormal representations R and S in Rr and Rs , respectively, with constants σR and σS . Let v = (v1 , . . . , vr ) be a vector in R and w = (w1 , . . . , ws ) be a vector in S. To the vertex in G × H corresponding to the pair (v, w) we associate the vector vwT := (v1 w1 , . . . , v1 ws , v2 w1 , . . . , v2 ws , . . . , vr w1 , . . . , vr ws ) ∈ Rrs. It is immediately checked that R × S := {vw T : v ∈ R, w ∈ S} is an orthonormal representation of G × H with constant σR σS . Hence by (6) we obtain μR×S (G × H) = μR (G)μS (H). For Gn = G × . . . × G and the representation T with constant σT this means μT n (Gn ) = μT (G)n = σTn and by (7) we obtain α(Gn ) ≤ σT−n ,

 n

α(Gn ) ≤ σT−1 .

Taking all things together we have thus completed Lovász’ argument:

Theorem. Whenever T = {v(1) , . . . , v (m) } is an orthonormal representation of G with constant σT , then Θ(G) ≤

1 . σT

(8)

1 T and hence Looking at the Lovász umbrella, we have u = (0, 0, h= √ 4 ) 5 √ 1 (i) 2 √ σ = v , u = h = 5 , which yields Θ(C5 ) ≤ 5. Thus Shannon’s problem is solved.

“Umbrellas with five ribs”

248

Communicating without errors

⎛ ⎜ ⎜ A=⎜ ⎜ ⎝

0 1 0 0 1

1 0 1 0 0

0 1 0 1 0

0 0 1 0 1

1 0 0 1 0

⎞ ⎟ ⎟ ⎟ ⎟ ⎠

The adjacency matrix for the 5-cycle C5

Let us carry our discussion a little further. We see from (8) that the larger σT is for a representation of G, the better a bound for Θ(G) we will get. Here is a method that gives us an orthonormal representation for any graph G. To G = (V, E) we associate the adjacency matrix A = (aij ), which is defined as follows: Let V = {v1 , . . . , vm }, then we set  1 if vi vj ∈ E aij := 0 otherwise. A is a real symmetric matrix with 0’s in the main diagonal. Now we need two facts from linear algebra. First, as a symmetric matrix, A has m real eigenvalues λ1 ≥ λ2 ≥ . . . ≥ λm (some of which may be equal), and the sum of the eigenvalues equals the sum of the diagonal entries of A, that is, 0. Hence the smallest eigenvalue must be negative (except in the trivial case when G has no edges). Let p = |λm | = −λm be the absolute value of the smallest eigenvalue, and consider the matrix M := I +

1 A, p

where I denotes the (m × m)-identity matrix. This M has the eigenvalues 1+ λp1 ≥ 1+ λp2 ≥ . . . ≥ 1+ λpm = 0. Now we quote the second result (the principal axis theorem of linear algebra): If M = (mij ) is a real symmetric matrix with all eigenvalues ≥ 0, then there are vectors v (1) , . . . , v (m) ∈ Rs for s = rank(M ), such that mij = v (i) , v (j) 

(1 ≤ i, j ≤ m).

In particular, for M = I + 1p A we obtain v (i) , v (i)  = mii = 1 and

v (i) , v (j)  =

1 aij p

for all i for i = j.

Since aij = 0 whenever vi vj ∈ E, we see that the vectors v (1) , . . . , v (m) form indeed an orthonormal representation of G. Let us, finally, apply this construction to the m-cycles Cm for odd m ≥ 5. π Here one easily computes p = |λmin | = 2 cos m (see the box). Every row of the adjacency matrix contains two 1’s, implying that every row of the matrix M sums to 1 + 2p . For the representation {v(1) , . . . , v (m) } this means 1 2 v (i) , v (1) + . . . + v (m)  = 1 + = 1 + π p cos m and hence

1 π −1 (1 + (cos m ) ) = σ m for all i. We can therefore apply our main result (8) and conclude m Θ(Cm ) ≤ (for m ≥ 5 odd). π −1 1 + (cos m ) v (i) , u =

(9)

Communicating without errors

249

π Notice that because of cos m < 1 the bound (9) is better than the bound √ m Θ(Cm ) ≤ 2 we found before. Note further cos π5 = τ2 , where τ = 5+1 2 is the golden section. Hence for m = 5 we again obtain

Θ(C5 ) ≤

5 1+

√4 5+1

√ 5( 5 + 1) √ √ = 5. = 5+ 5

The orthonormal representation given by this construction is, of course, precisely the “Lovász umbrella.” 2 And what about C7 , C9 , and the other odd cycles? By considering α(Cm ), m−1 3 α(Cm ) and other small powers the lower bound 2 ≤ Θ(Cm ) can certainly be increased, but for no odd m ≥ 7 do the best known lower bounds agree with the upper bound given √ in (8). So, twenty years after Lovász’ marvelous proof of Θ(C5 ) = 5, these problems remain open and are considered very difficult — but after all we had this situation before.

The eigenvalues of Cm Look at the adjacency matrix A of the cycle Cm . To find the eigenvalues (and eigenvectors) we use the m-th roots of unity. These are 2πi given by 1, ζ, ζ 2 , . . . , ζ m−1 for ζ = e m — see the box on page 33. Let λ = ζ k be any of these roots, then we claim that (1, λ, λ2 , . . . , λm−1 )T is an eigenvector of A to the eigenvalue λ + λ−1 . In fact, by the set-up of A we find ⎛ ⎞ ⎛ ⎞ ⎞ ⎛ λ + λm−1 1 1 ⎜ λ ⎟ ⎜ λ2 + 1 ⎟ ⎟ ⎜ λ ⎜ 2 ⎟ ⎜ 3 ⎟ ⎟ ⎜ ⎜ λ ⎟ ⎜ λ + λ ⎟ ⎟ −1 ⎜ λ2 A⎜ ⎟=⎜ ⎟ = (λ + λ ) ⎜ ⎟. ⎜ .. ⎟ ⎜ ⎟ ⎟ ⎜ .. .. ⎝ . ⎠ ⎝ ⎠ ⎠ ⎝ . . m−1 m−2 m−1 λ λ 1 + λ Since the vectors (1, λ, . . . , λm−1 ) are independent (they form a socalled Vandermonde matrix) we conclude that for odd m ζ k + ζ −k

=

[(cos(2kπ/m) + i sin(2kπ/m)] + [cos(2kπ/m) − i sin(2kπ/m)]

=

2 cos(2kπ/m)

(0 ≤ k ≤

m−1 2 )

are all the eigenvalues of A. Now the cosine is a decreasing function, and so  (m − 1)π  π 2 cos = −2 cos m m is the smallest eigenvalue of A.

For example, for m = 7 all we know is √ 7 4 108 ≤ Θ(C7 ) ≤ , 1 + (cos π7 )−1 which is 3.2237 ≤ Θ(C7 ) ≤ 3.3177.

250

Communicating without errors

References [1] V. C HVÁTAL : Linear Programming, Freeman, New York 1983. [2] W. H AEMERS : Eigenvalue methods, in: “Packing and Covering in Combinatorics” (A. Schrijver, ed.), Math. Centre Tracts 106 (1979), 15-38. [3] L. L OVÁSZ : On the Shannon capacity of a graph, IEEE Trans. Information Theory 25 (1979), 1-7. [4] C. E. S HANNON : The zero-error capacity of a noisy channel, IRE Trans. Information Theory 3 (1956), 3-15.

The chromatic number of Kneser graphs

In 1955 the number theorist Martin Kneser posed a seemingly innocuous problem that became one of the great challenges in graph theory until a brilliant and totally unexpected solution, using the “Borsuk–Ulam theorem” from topology, was found by László Lovász twenty-three years later. It happens often in mathematics that once a proof for a long-standing problem is found, a shorter one quickly follows, and so it was in this case. Within weeks Imre Bárány showed how to combine the Borsuk–Ulam theorem with another known result to elegantly settle Kneser’s conjecture. Then in 2002 Joshua Greene, an undergraduate student, simplified Bárány’s argument even further, and it is his version of the proof that we present here. But let us start at the beginning. Consider the following graph K(n, k), now called Kneser graph, for integers n ≥ k ≥ 1. The vertex-set

 V (n, k) is the family of k-subsets of {1, . . . , n}, thus |V (n, k)| = nk . Two such k-sets A and B are adjacent if they are disjoint, A ∩ B = ∅. If n < 2k, then any two k-sets intersect, resulting in the uninteresting case where K(n, k) has no edges. So we assume from now on that n ≥ 2k. Kneser graphs provide an interesting link between graph theory and finite sets. Consider, e.g., the independence number α(K(n, k)), that is, we ask how large a family of pairwise intersecting k-sets can be. The answer given

isn−1  by the Erd˝os–Ko–Rado theorem of Chapter 27: α(K(n, k)) = k−1 . We can similarly study other interesting parameters of this graph family, and Kneser picked out the most challenging one: the chromatic number χ(K(n, k)). We recall from previous chapters that a (vertex) coloring of a graph G is a mapping c : V → {1, . . . , m} such that adjacent vertices are colored differently. The chromatic number χ(G) is then the minimum number of colors that is sufficient for a coloring of V . In other words, we want to present the vertex set V as a disjoint union of as few color classes as possible, V = V1 ∪˙ · · · ∪˙ Vχ(G) , such that each set Vi is edgeless. For the graphs K(n, k) this asks for a partition V (n, k) = V1 ∪˙ · · · ∪˙ Vχ , where every Vi is an intersecting family of k-sets. Since we assume that n ≥ 2k, we write from now on n = 2k + d, k ≥ 1, d ≥ 0. Here is a simple coloring of K(n, k) that uses d + 2 colors: For i = 1, 2, . . . , d + 1, let Vi consist of all k-sets that have i as smallest element. The remaining k-sets are contained in the set {d + 2, d + 3, . . . , 2k + d}, which has only 2k − 1 elements. Hence they all intersect, and we can use color d + 2 for all of them.

M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_38, © Springer-Verlag Berlin Heidelberg 2013

Chapter 38

{1, 2}

{3, 5}

{4, 5}

{3, 4} {2, 5}

{1, 3}

{2, 4}

{1, 4}

{2, 3} {1, 5} The Kneser graph K(5, 2) is the famous Petersen graph.

This implies that χ(K(n, k)) ≥

|V | α

=

(nk) = (n−1 k−1 )

n . k

1

3

3

3 2

1 1

2

2 1 The 3-coloring of the Petersen graph.

252

The chromatic number of Kneser graphs So we have χ(K(2k + d, k)) ≤ d + 2, and Kneser’s challenge was to show that this is the right number. Kneser’s conjecture: We have χ(K(2k + d, k)) = d + 2.

For d = 0, K(2k, k) consists of disjoint edges, one for every pair of complementary k-sets. Hence χ(K(2k, k)) = 2, in accordance with the conjecture.

Probably anybody’s first crack at the proof would be to try induction on k and d. Indeed, the starting cases k = 1 and d = 0, 1 are easy, but the induction step from k to k + 1 (or d to d + 1) does not seem to work. So let us instead reformulate the conjecture as an existence problem: If the family of k-sets of {1, 2, . . . , 2k + d} is partitioned into d + 1 classes, V (n, k) = V1 ∪˙ · · · ∪˙ Vd+1 , then for some i, Vi contains a pair A, B of disjoint k-sets. Lovász’ brilliant insight was that at the (topological) heart of the problem lies a famous theorem about the d-dimensional unit sphere S d in Rd+1 , S d = {x ∈ Rd+1 : |x| = 1}. The Borsuk–Ulam theorem. For every continuous map f : S d → Rd from d-sphere to d-space, there are antipodal points x∗ , −x∗ that are mapped to the same point f (x∗ ) = f (−x∗ ). This result is one of the cornerstones of topology; it first appears in Borsuk’s famous 1933 paper. We sketch a proof in the appendix; for the full proof we refer to Section 2.2 in Matoušek’s wonderful book “Using the Borsuk–Ulam theorem”, whose very title demonstrates the power and range of the result. Indeed, there are many equivalent formulations, which underline the central position of the theorem. We will employ a version that can be traced back to a book by Lyusternik–Shnirel’man from 1930, which even predates Borsuk. Theorem. If the d-sphere S d is covered by d + 1 sets, S d = U1 ∪ · · · ∪ Ud ∪ Ud+1 , such that each of the first d sets U1 , . . . , Ud is either open or closed, then one of the d + 1 sets contains a pair of antipodal points x∗ , −x∗ . The case when all d+1 sets are closed is due to Lyusternik and Shnirel’man. The case when all d+1 sets are open is equally common, and also called the Lyusternik–Shnirel’man theorem. Greene’s insight was that the theorem is also true if each of the d + 1 sets is either open or closed. As you will see, we don’t even need that: No such assumption is needed for Ud+1 . For the proof of Kneser’s conjecture, we only need the case when U1 , . . . , Ud are open.

The chromatic number of Kneser graphs

253

 Proof of the Lyusternik–Shnirel’man theorem using Borsuk–Ulam. Let a covering S d = U1 ∪ · · · ∪ Ud ∪ Ud+1 be given as specified, and assume that there are no antipodal points in any of the sets Ui . We define a map f : S d → Rd by 

f (x) := δ(x, U1 ), δ(x, U2 ), . . . , δ(x, Ud ) . Here δ(x, Ui ) denotes the distance of x from Ui . Since this is a continuous function in x, the map f is continuous. Thus the Borsuk–Ulam theorem tells us that there are antipodal points x∗ , −x∗ with f (x∗ ) = f (−x∗ ). Since Ud+1 does not contain antipodes, we get that at least one of x∗ and −x∗ must be contained in one of the sets Ui , say in Uk (k ≤ d). After exchanging x∗ with −x∗ if necessary, we may assume that x∗ ∈ Uk . In particular this yields δ(x∗ , Uk ) = 0, and from f (x∗ ) = f (−x∗ ) we get that δ(−x∗ , Uk ) = 0 as well. If Uk is closed, then δ(−x∗ , Uk ) = 0 implies that −x∗ ∈ Uk , and we arrive at the contradiction that Uk contains a pair of antipodal points. If Uk is open, then δ(−x∗ , Uk ) = 0 implies that −x∗ lies in Uk , the closure of Uk . The set Uk , in turn, is contained in S d \(−Uk ), since this is a closed subset of S d that contains Uk . But this means that −x∗ lies in S d \(−Uk ), so it cannot lie in −Uk , and x∗ cannot lie in Uk , a contradiction. 

The closure of Uk is the smallest closed set that contains Uk (that is, the intersection of all closed sets containing Uk ).

As the second ingredient for his proof, Imre Bárány used another existence result about the sphere S d . Gale’s Theorem. There is an arrangement of 2k + d points on S d such that every open hemisphere contains at least k of these points. David Gale discovered his theorem in 1956 in the context of polytopes with many faces. He presented a complicated induction proof, but today, with hindsight, we can quite easily exhibit such a set and verify its properties. Armed with these results it is just a short step to settle Kneser’s problem, but as Greene showed we can do even better: We don’t even need Gale’s result. It suffices to take any arrangement of 2k + d points on S d+1 in general position, meaning that no d + 2 of the points lie on a hyperplane through the center of the sphere. Clearly, for d ≥ 0 this can be done.  Proof of the Kneser conjecture. For our ground set let us take 2k + d points in general position on the sphere S d+1 . Suppose the set V (n, k) of all k-subsets of this set is partitioned into d + 1 classes, V (n, k) = V1 ∪˙ · · · ∪˙ Vd+1 . We have to find a pair of disjoint k-sets A and B that belong to the same class Vi . For i = 1, . . . , d + 1 we set Oi = {x ∈ S

d+1

: the open hemisphere Hx with pole x contains a k-set from Vi } .

Clearly, each Oi is an open set. Together, the open sets Oi and the closed set C = S d+1 \ (O1 ∪ · · · ∪ Od+1 ) cover S d+1 . Invoking Lyusternik– Shnirel’man we know that one of these sets contains antipodal points x∗

x Hx An open hemisphere in S 2

254

The chromatic number of Kneser graphs and −x∗ . This set cannot be C! Indeed, if x∗ , −x∗ are in C, then by the definition of the Oi ’s, the hemispheres Hx∗ and H−x∗ would contain fewer than k points. This means that at least d + 2 points would be on the equator H x∗ ∩ H −x∗ with respect to the north pole x∗ , that is, on a hyperplane through the origin. But this cannot be since the points are in general position. Hence some Oi contains a pair x∗ , −x∗ , so there exist k-sets A and B both in class Vi , with A ⊆ Hx∗ and B ⊆ H−x∗ . x∗ A −→

B −→

Hx ∗

−x∗

H−x∗

But since we are talking about open hemispheres, Hx∗ and H−x∗ are disjoint, hence A and B are disjoint, and this is the whole proof.  The reader may wonder whether sophisticated results such as the theorem of Borsuk–Ulam are really necessary to prove a statement about finite sets. Indeed, a beautiful combinatorial argument has recently been found by Jiˇrí Matoušek — but on closer inspection it has a distinct, albeit discrete, topological flavor.

Appendix: A proof sketch for the Borsuk–Ulam theorem For any generic map (also known as general position map) from a compact d-dimensional space to a d-dimensional space, any point in the image has only a finite number of pre-images. For a generic map from a (d + 1)dimensional space to a d-dimensional space, we expect every point in the image to have a 1-dimensional pre-image, that is, a collection of curves. Both in the case of smooth maps, and in the setting of piecewise-linear maps, one quite easily proves one can deform any map to a nearby generic map. For the Borsuk–Ulam theorem, the idea is to show that every generic map S d → Rd identifies an odd (in particular, finite and nonzero) number of antipodal pairs. If f did not identify any antipodal pair, then it would be arbitrarily close to a generic map f' without any such identification. Now consider the projection π : S d → Rd that just deletes the last coordinate; this map identifies the “north pole” ed+1 of the d-sphere with the “south pole” −ed+1 . For any given map f : S d → Rd we construct a continuous deformation from π to f , that is, we interpolate between these two

The chromatic number of Kneser graphs

255

maps (linearly, for example), to obtain a continuous map

Sd

F : S × [0, 1] −→ R , d

d

t=1

with F (x, 0) = π(x) and F (x, 1) = f (x) for all x ∈ S . (Such a map is known as a homotopy.) Now we perturb F carefully into a generic map F' : S d ×[0, 1] → Rd , which again we may assume to be smooth, or piecewise-linear on a fine triangulation of S d × [0, 1]. If this perturbation is “small enough” and performed carefully, then the perturbed version of the projection π '(x) := F' (x, 0) should still identify the two antipodal points ±ed+1 and no others. If F' is sufficiently generic, then the points in S d × [0, 1] given by 7 8 M := (x, t) ∈ S d × [0, 1] : F' (−x, t) = F' (x, t)

f

d

according to the implicit function theorem (smooth or piecewise-linear version) form a collection of paths and of closed curves. Clearly this collection is symmetric, that is, (−x, t) ∈ M if and only if (x, t) ∈ M . The paths in M can have endpoints only at the boundary of S d × [0, 1], that is, at t = 0 and at t = 1. The only ends at t = 0, however, are at (±ed+1 , 0), and the two paths that start at these two points are symmetric copies of each other, so they are disjoint, and they can end only at t = 1. This proves that there are solutions for F' (−x, t) = F' (x, t) at t = 1, and hence for f (−x) = f (x). 

References [1] I. BÁRÁNY: A short proof of Kneser’s conjecture, J. Combinatorial Theory, Ser. B 25 (1978), 325-326. [2] K. B ORSUK : Drei Sätze über die n-dimensionale Sphäre, Fundamenta Math. 20 (1933), 177-190. [3] D. G ALE : Neighboring vertices on a convex polyhedron, in: “Linear Inequalities and Related Systems” (H. W. Kuhn, A. W. Tucker, eds.), Princeton University Press, Princeton 1956, 255-263. [4] J. E. G REENE : A new short proof of Kneser’s conjecture, American Math. Monthly 109 (2002), 918-920. [5] M. K NESER : Aufgabe 360, Jahresbericht der Deutschen MathematikerVereinigung 58 (1955), 27. [6] L. L OVÁSZ : Kneser’s conjecture, chromatic number, and homotopy, J. Combinatorial Theory, Ser. B 25 (1978), 319-324. [7] L. LYUSTERNIK & S. S HNIREL’ MAN : Topological Methods in Variational Problems (in Russian), Issledowatelski˘ı Institute Matematiki i Mechaniki pri O. M. G. U., Moscow, 1930. [8] J. M ATOUŠEK : Using the Borsuk–Ulam Theorem. Lectures on Topological Methods in Combinatorics and Geometry, Universitext, Springer-Verlag, Berlin 2003. [9] J. M ATOUŠEK : A combinatorial proof of Kneser’s conjecture, Combinatorica 24 (2004), 163-170.

S d × [0, 1]

F

Rd π

t=0

Sd

(S d , 1) t=1

f F (ed+1 , 0)

Rd π

t=0

(−ed+1 , 0)

(S d , 0)

Of friends and politicians

Chapter 39

It is not known who first raised the following problem or who gave it its human touch. Here it is: Suppose in a group of people we have the situation that any pair of persons have precisely one common friend. Then there is always a person (the “politician”) who is everybody’s friend. In the mathematical jargon this is called the friendship theorem. Before tackling the proof let us rephrase the problem in graph-theoretic terms. We interpret the people as the set of vertices V and join two vertices by an edge if the corresponding people are friends. We tacitly assume that friendship is always two-ways, that is, if u is a friend of v, then v is also a friend of u, and further that nobody is his or her own friend. Thus the theorem takes on the following form:

“A politician’s smile”

Theorem. Suppose that G is a finite graph in which any two vertices have precisely one common neighbor. Then there is a vertex which is adjacent to all other vertices. Note that there are finite graphs with this property; see the figure, where u is the politician. However, these “windmill graphs” also turn out to be the only graphs with the desired property. Indeed, it is not hard to verify that in the presence of a politician only the windmill graphs are possible. Surprisingly, the friendship theorem does not hold for infinite graphs! Indeed, for an inductive construction of a counterexample one may start for example with a 5-cycle, and repeatedly add common neighbors for all pairs of vertices in the graph that don’t have one, yet. This leads to a (countably) infinite friendship graph without a politician. Several proofs of the friendship theorem exist, but the first proof, given by Paul Erd˝os, Alfred Rényi and Vera Sós, is still the most accomplished.  Proof. Suppose the assertion is false, and G is a counterexample, that is, no vertex of G is adjacent to all other vertices. To derive a contradiction we proceed in two steps. The first part is combinatorics, and the second part is linear algebra. (1) We claim that G is a regular graph, that is, d(u) = d(v) for any u, v ∈ V. Note first that the condition of the theorem implies that there are no cycles of length 4 in G. Let us call this the C4 -condition. M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_39, © Springer-Verlag Berlin Heidelberg 2013

.. .

u

A windmill graph

u

v

258

Of friends and politicians

u

w1

w2 w3

...

wk

... z2 v

z3 . . . zk

We first prove that any two non-adjacent vertices u and v have equal degree d(u) = d(v). Suppose d(u) = k, where w1 , . . . , wk are the neighbors of u. Exactly one of the wi , say w2 , is adjacent to v, and w2 adjacent to exactly one of the other wi ’s, say w1 , so that we have the situation of the figure to the left. The vertex v has with w1 the common neighbor w2 , and with wi (i ≥ 2) a common neighbor zi (i ≥ 2). By the C4 -condition, all these zi must be distinct. We conclude d(v) ≥ k = d(u), and thus d(u) = d(v) = k by symmetry. To finish the proof of (1), observe that any vertex different from w2 is not adjacent to either u or v, and hence has degree k, by what we already proved. But since w2 also has a non-neighbor, it has degree k as well, and thus G is k-regular. Summing over the degrees of the k neighbors of u we get k 2 . Since every vertex (except u) has exactly one common neighbor with u, we have counted every vertex once, except for u, which was counted k times. So the total number of vertices of G is n = k 2 − k + 1.

(1)

(2) The rest of the proof is a beautiful application of some standard results of linear algebra. Note first that k must be greater than 2, since for k ≤ 2 only G = K1 and G = K3 are possible by (1), both of which are trivial windmill graphs. Consider the adjacency matrix A = (aij ), as defined on page 248. By part (1), any row has exactly k 1’s, and by the condition of the theorem, for any two rows there is exactly one column where they both have a 1. Note further that the main diagonal consists of 0’s. Hence we have ⎞ ⎛ k 1 ... 1 ⎜ 1 k 1 ⎟ ⎟ ⎜ A2 = ⎜ . .. ⎟ = (k − 1) I + J , . . . ⎝ . . . ⎠ 1 ... 1 k where I is the identity matrix, and J the matrix of all 1’s. It is immediately checked that J has the eigenvalues n (of multiplicity 1) and 0 (of multiplicity n − 1). It follows that A2 has the eigenvalues k − 1 + n = k 2 (of multiplicity 1) and k − 1 (of multiplicity n − 1). Since A is symmetric and hence diagonalizable, √ we conclude that A has the eigenvalues k (of multiplicity 1) and ± k − 1. Suppose √ √ r of the eigenvalues are equal to k − 1 and s of them are equal to − k − 1, with r + s = n − 1. Now we are almost home. Since the sum of the eigenvalues of A equals the trace (which is 0), we find √ √ k + r k − 1 − s k − 1 = 0, and, in particular, r = s, and √ k−1 =

k . s−r

Of friends and politicians √ Now if the square root m of a natural number m is rational, then it is an integer! An elegant proof for this was presented by Dedekind in 1858: Let √ √ n0 be the smallest natural number with n m ∈ N. If m ∈  N, then there 0 √ √ exists  ∈ N with 0 < m −  < 1. Setting n := n ( m − ), we find 1 0 √ √ √ √ n1 ∈ N and n1 m = n0 ( m − ) m = n0 m − (n0 m) ∈ N. With n1 < n0 this yields a contradiction to the choice of n0 . √ Returning to our equation, let us set h = k − 1 ∈ N, then h(s − r) = k = h2 + 1. Since h divides h2 + 1 and h2 , we find that h must be equal to 1, and thus k = 2, which we have already excluded. So we have arrived at a contradiction, and the proof is complete.  However, the story is not quite over. Let us rephrase our theorem in the following way: Suppose G is a graph with the property that between any two vertices there is exactly one path of length 2. Clearly, this is an equivalent formulation of the friendship condition. Our theorem then says that the only such graphs are the windmill graphs. But what if we consider paths of length more than 2? A conjecture of Anton Kotzig asserts that the analogous situation is impossible. Kotzig’s Conjecture. Let  > 2. Then there are no finite graphs with the property that between any two vertices there is precisely one path of length . Kotzig himself verified his conjecture for  ≤ 8. In [3] his conjecture is proved up to  = 20, and A. Kostochka has told us recently that it is now verified for all  ≤ 33. A general proof, however, seems to be out of reach . . .

References ˝ , A. R ÉNYI & V. S ÓS : On a problem of graph theory, Studia Sci. [1] P. E RD OS Math. 1 (1966), 215-235.

[2] A. KOTZIG : Regularly k-path connected graphs, Congressus Numerantium 40 (1983), 137-141. [3] A. KOSTOCHKA : The nonexistence of certain generalized friendship graphs, in: “Combinatorics” (Eger, 1987), Colloq. Math. Soc. János Bolyai 52, NorthHolland, Amsterdam 1988, 341-356.

259

Probability makes counting (sometimes) easy

Chapter 40

Just as we started this book with the first papers of Paul Erd˝os in number theory, we close it by discussing what will possibly be considered his most lasting legacy — the introduction, together with Alfred Rényi, of the probabilistic method. Stated in the simplest way it says: If, in a given set of objects, the probability that an object does not have a certain property is less than 1, then there must exist an object with this property. Thus we have an existence result. It may be (and often is) very difficult to find this object, but we know that it exists. We present here three examples (of increasing sophistication) of this probabilistic method due to Erd˝os, and end with a particularly elegant recent application. As a warm-up, consider a family F of subsets Ai , all of size d ≥ 2, of a finite ground-set X. We say that F is 2-colorable if there exists a coloring of X with two colors such that in every set Ai both colors appear. It is immediate that not every family can be colored in this way. As an example, take all subsets of size d of a (2d − 1)-set X. Then no matter how we 2-color X, there must be d elements which are colored alike. On the other hand, it is equally clear that every subfamily of a 2-colorable family of d-sets is itself 2-colorable. Hence we are interested in the smallest number m = m(d) for which a family with m sets exists which is not 2-colorable. Phrased differently, m(d) is the largest number which guarantees that every family with less than m(d) sets is 2-colorable. Theorem 1. Every family of at most 2d−1 d-sets is 2-colorable, that is, m(d) > 2d−1 .  Proof. Suppose F is a family of d-sets with at most 2d−1 sets. Color X randomly with two colors, all colorings being equally likely. For each set A ∈ F let EA be the event that all elements of A are colored alike. Since there are precisely two such colorings, we have d−1

Prob(EA ) = ( 12 )

,

and hence with m = |F | ≤ 2 (note that the events EA are not disjoint)   d−1 EA ) < Prob(EA ) = m ( 12 ) ≤ 1. Prob( d−1

A∈F

A∈F

We conclude that there exists some 2-coloring of X without a unicolored d-set from F , and this is just our condition of 2-colorability.  M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6_40, © Springer-Verlag Berlin Heidelberg 2013

A 2-colored family of 3-sets

262

Probability makes counting (sometimes) easy An upper bound for m(d), roughly equal to d2 2d , was also established by Erd˝os, again using the probabilistic method, this time taking random sets and a fixed coloring. As for exact values, only the first two m(2) = 3, m(3) = 7 are known. Of course, m(2) = 3 is realized by the graph K3 , while the Fano configuration yields m(3) ≤ 7. Here F consists of the seven 3-sets of the figure (including the circle set {4, 5, 6}). The reader may find it fun to show that F needs 3 colors. To prove that all families of six 3-sets are 2-colorable, and hence m(3) = 7, requires a little more care.

3

5

1

7

4

2

6

Our next example is the classic in the field — Ramsey numbers. Consider the complete graph KN on N vertices. We say that KN has property (m, n) if, no matter how we color the edges of KN red and blue, there is always a complete subgraph on m vertices with all edges colored red or a complete subgraph on n vertices with all edges colored blue. It is clear that if KN has property (m, n), then so does every Ks with s ≥ N . So, as in the first example, we ask for the smallest number N (if it exists) with this property — and this is the Ramsey number R(m, n). As a start, we certainly have R(m, 2) = m because either all of the edges of Km are red or there is a blue edge, resulting in a blue K2 . By symmetry, we have R(2, n) = n. Now, suppose R(m − 1, n) and R(m, n − 1) exist. We then prove that R(m, n) exists and that R(m, n) ≤ R(m − 1, n) + R(m, n − 1).

red edges

A

v

blue edges

B

(1)

Suppose N = R(m − 1, n) + R(m, n − 1), and consider an arbitrary redblue coloring of KN . For a vertex v, let A be the set of vertices joined to v by a red edge, and B the vertices joined by a blue edge. Since |A| + |B| = N − 1, we find that either |A| ≥ R(m − 1, n) or |B| ≥ R(m, n − 1). Suppose |A| ≥ R(m − 1, n), the other case being analogous. Then by the definition of R(m − 1, n), there either exists in A a subset AR of size m − 1 all of whose edges are colored red which together with v yields a red Km , or there is a subset AB of size n with all edges colored blue. We infer that KN satisfies the (m, n)-property and Claim (1) follows. Combining (1) with the starting values R(m, 2) = m and R(2, n) = n, we obtain from the familiar recursion for binomial coefficients   m+n−2 R(m, n) ≤ , (2) m−1 and, in particular  R(k, k) ≤

2k − 2 k−1



    2k − 3 2k − 3 = + ≤ 22k−3 . k−1 k−2

Now what we are really interested in is a lower bound for R(k, k). This amounts to proving for an as-large-as-possible N < R(k, k) that there exists a coloring of the edges such that no red or blue Kk results. And this is where the probabilistic method comes into play.

Probability makes counting (sometimes) easy

263

Theorem 2. For all k ≥ 2, the following lower bound holds for the Ramsey numbers: k

R(k, k) ≥ 2 2 .  Proof. We have R(2, 2) = 2. From (2) we know R(3, 3) ≤ 6, and the pentagon colored as in the figure shows R(3, 3) = 6. k

Now let us assume k ≥ 4. Suppose N < 2 2 , and consider all red-blue colorings, where we color each edge independently red or blue with probaN bility 12 . Thus all colorings are equally likely with probability 2−( 2 ) . Let A be a set of vertices of size k. The probability of the event AR that the edges k in A are all colored red is then 2−(2) . Hence it follows that the probability pR for some k-set to be colored all red is bounded by pR



 = Prob AR ≤



|A|=k

|A|=k

  N −(k2 ) . Prob(AR ) = 2 k

 k Nk Now with N < 2 2 and k ≥ 4, using N k ≤ 2k−1 for k ≥ 2 (see page 12), we have   k k k2 k N −(k2) Nk 1 2 ≤ k−1 2−(2) < 2 2 −(2)−k+1 = 2− 2 +1 ≤ . k 2 2 Hence pR < 12 , and by symmetry pB < 12 for the probability of some k vertices with all edges between them colored blue. We conclude that k pR + pB < 1 for N < 2 2 , so there must be a coloring with no red or blue Kk , which means that KN does not have property (k, k).  Of course, there is quite a gap between the lower and the upper bound for R(k, k). Still, as simple as this Book Proof is, no lower bound with a better exponent has been found for general k in the more than 50 years since Erd˝os’ result. In fact, no one has been able to prove a lower bound of the 1 form R(k, k) > 2( 2 +ε)k nor an upper bound of the form R(k, k) < 2(2−ε)k for a fixed ε > 0. Our third result is another beautiful illustration of the probabilistic method. Consider a graph G on n vertices and its chromatic number χ(G). If χ(G) is high, that is, if we need many colors, then we might suspect that G contains a large complete subgraph. However, this is far from the truth. Already in the fourties Blanche Descartes constructed graphs with arbitrarily high chromatic number and no triangles, that is, with every cycle having length at least 4, and so did several others (see the box on the next page). However, in these examples there were many cycles of length 4. Can we do even better? Can we stipulate that there are no cycles of small length and still have arbitrarily high chromatic number? Yes we can! To make matters precise, let us call the length of a shortest cycle in G the girth γ(G) of G; then we have the following theorem, first proved by Paul Erd˝os.

red blue

264

Probability makes counting (sometimes) easy

Triangle-free graphs with high chromatic number G3 :

Here is a sequence of triangle-free graphs G3 , G4 , . . . with χ(Gn ) = n.

G4 :

Constructing the Mycielski graph

Start with G3 = C5 , the 5-cycle; thus χ(G3 ) = 3. Suppose we have already constructed Gn on the vertex set V . The new graph Gn+1 has the vertex set V ∪ V ∪ {z}, where the vertices v ∈ V correspond bijectively to v ∈ V , and z is a single other vertex. The edges of Gn+1 fall into 3 classes: First, we take all edges of Gn ; secondly every vertex v is joined to precisely the neighbors of v in Gn ; thirdly z is joined to all v ∈ V . Hence from G3 = C5 we obtain as G4 the so-called Mycielski graph. Clearly, Gn+1 is again triangle-free. To prove χ(Gn+1 ) = n + 1 we use induction on n. Take any n-coloring of Gn and consider a color class C. There must exist a vertex v ∈ C which is adjacent to at least one vertex of every other color class; otherwise we could distribute the vertices of C onto the n − 1 other color classes, resulting in χ(Gn ) ≤ n − 1. But now it is clear that v (the vertex in V corresponding to v) must receive the same color as v in this n-coloring. So, all n colors appear in V , and we need a new color for z.

Theorem 3. For every k ≥ 2, there exists a graph G with chromatic number χ(G) > k and girth γ(G) > k. The strategy is similar to that of the previous proofs: We consider a certain probability space on graphs and go on to show that the probability for χ(G) ≤ k is smaller than 12 , and similarly the probability for γ(G) ≤ k is smaller than 12 . Consequently, there must exist a graph with the desired properties.  Proof. Let V = {v1 , v2 , . . . , vn } be the vertex set, and p a fixed number between 0 and 1, to be carefully chosen later. Our probability space G(n, p) consists of all graphs on V where the individual edges appear with probability p, independently of each other. In other words, we are talking about a Bernoulli experiment where we throw in each edge with probability p. As an example, the probability Prob(Kn ) for the complete graph n n is Prob(Kn ) = p( 2 ) . In general, we have Prob(H) = pm (1 − p)( 2 )−m if the graph H on V has precisely m edges. Let us first look at the chromatic number χ(G). By α = α(G) we denote the independence number, that is, the size of a largest independent set in G. Since in a coloring with χ = χ(G) colors all color classes are independent (and hence of size ≤ α), we infer χα ≥ n. Therefore if α is small as compared to n, then χ must be large, which is what we want. Suppose 2 ≤ r ≤ n. The probability that a fixed r-set in V is independent

Probability makes counting (sometimes) easy

265

r is (1 − p)(2) , and we conclude by the same argument as in Theorem 2   r n Prob(α ≥ r) ≤ (1 − p)(2) r r r−1 ≤ nr (1 − p)(2) = (n(1 − p) 2 )r ≤ (ne−p(r−1)/2 )r ,

since 1 − p ≤ e−p for all p. k

Given any fixed k > 0 we now choose p := n− k+1 , and proceed to show that for n large enough,  n 1 Prob α ≥ < . (3) 2k 2 1

1

Indeed, since n k+1 grows faster than log n, we have n k+1 ≥ 6k log n n for large enough n, and thus p ≥ 6k logn n . For r :=  2k  this gives pr ≥ 3 log n, and thus ne−p(r−1)/2 = ne−

pr 2

p

3

1

1

1

1

e 2 ≤ ne− 2 log n e 2 = n− 2 e 2 = ( ne ) 2 ,

which converges to 0 as n goes to infinity. Hence (3) holds for all n ≥ n1 . Now we look at the second parameter, γ(G). For the given k we want to show that there are not too many cycles of length ≤ k. Let i be between 3 and k, and A ⊆ V a fixed i-set. The number of possible i-cycles on A is clearly the number of cyclic permutations of A divided by 2 (since we may traverse the cycle in either direction), and thus equal to (i−1)! 2 . The total

n (i−1)! number of possible i-cycles is therefore i 2 , and every such cycle C appears with probability pi . Let X be the random variable which counts the number of cycles of length ≤ k. In order to estimate X we use two simple but beautiful tools. The first is linearity of expectation, and the second is Markov’s inequality for nonnegative random variables, which says Prob(X ≥ a) ≤

EX , a

where EX is the expected value of X. See the appendix to Chapter 15 for both tools. Let XC be the indicator random variable of the cycle C of, say, length i. That is, we set XC = 1 or 0 depending on whether C appears in the graph i or not; hence EXC = p

. Since X counts the number of all cycles of length ≤ k we have X = XC , and hence by linearity EX =

k    n (i − 1)! i=3

i

2

1 i i 1 np ≤ (k − 2)nk pk 2 i=3 2 k

pi ≤

1

where the last inequality holds because of np = n k+1 ≥ 1. Applying now Markov’s inequality with a = n2 , we obtain Prob(X ≥

n 2)

≤

1 EX (np)k ≤ (k − 2) = (k − 2)n− k+1 . n/2 n

266

Probability makes counting (sometimes) easy Since the right-hand side goes to 0 with n going to infinity, we infer that p(X ≥ n2 ) < 12 for n ≥ n2 . Now we are almost home. Our analysis tells us that for n ≥ max(n1 , n2 ) n there exists a graph H on n vertices with α(H) < 2k and fewer than n2 cycles of length ≤ k. Delete one vertex from each of these cycles, and let G be the resulting graph. Then γ(G) > k holds at any rate. Since G n contains more than n2 vertices and satisfies α(G) ≤ α(H) < 2k , we find χ(G) ≥

n n n/2 ≥ > = k, α(G) 2α(H) n/k

and the proof is finished.



Explicit constructions of graphs with high girth and chromatic number (of huge size) are known. (In contrast, one does not know how to construct red/blue colorings with no large monochromatic cliques, whose existence is given by Theorem 2.) What remains striking about the Erd˝os proof is that it proves the existence of relatively small graphs with high chromatic number and girth. To end our excursion into the probabilistic world let us discuss an important result in geometric graph theory (which again goes back to Paul Erd˝os) whose stunning Book Proof is of recent vintage. Consider a simple graph G = G(V, E) with n vertices and m edges. We want to embed G into the plane just as we did for planar graphs. Now, we know from Chapter 12 — as a consequence of Euler’s formula — that a simple planar graph G has at most 3n − 6 edges. Hence if m is greater than 3n − 6, there must be crossings of edges. The crossing number cr(G) is then naturally defined: It is the smallest number of crossings among all drawings of G, where crossings of more than two edges in one point are not allowed. Thus cr(G) = 0 if and only if G is planar. In such a minimal drawing the following three situations are ruled out: • No edge can cross itself. • Edges with a common endvertex cannot cross. • No two edges cross twice. This is because in either of these cases, we can construct a different drawing of the same graph with fewer crossings, using the operations that are indicated in our figure. So, from now on we assume that any drawing observes these rules. Suppose that G is drawn in the plane with cr(G) crossings. We can immediately derive a lower bound on the number of crossings. Consider the following graph H: The vertices of H are those of G together with all crossing points, and the edges are all pieces of the original edges as we go along from crossing point to crossing point. The new graph H is now plane and simple (this follows from our three assumptions!). The number of vertices in H is n + cr(G) and the number

Probability makes counting (sometimes) easy

267

of edges is m + 2cr(G), since every new vertex has degree 4. Invoking the bound on the number of edges for plane graphs we thus find m + 2 cr(G) ≤ 3(n + cr(G)) − 6, that is, cr(G) ≥ m − 3n + 6.

(4)

As an example, for the complete graph K6 we compute cr(K6 ) ≥ 15 − 18 + 6 = 3 and, in fact, there is an drawing with just 3 crossings. The bound (4) is good enough when m is linear in n, but when m is larger compared to n, then the picture changes, and this is our theorem. Theorem 4. Let G be a simple graph with n vertices and m edges, where m ≥ 4n. Then 1 m3 cr(G) ≥ . 64 n2 The history of this result, called the crossing lemma, is quite interesting. 1 It was conjectured by Erd˝os and Guy in 1973 (with 64 replaced by some 1 constant c). The first proofs were given by Leighton in 1982 (with 100 in1 stead of 64 ) and independently by Ajtai, Chvátal, Newborn and Szemerédi. The crossing lemma was hardly known (in fact, many people thought of it as a conjecture long after the original proofs), until László Székely demonstrated its usefulness in a beautiful paper, applying it to a variety of hitherto hard geometric extremal problems. The proof which we now present arose from e-mail conversations between Bernard Chazelle, Micha Sharir and Emo Welzl, and it belongs without doubt in The Book.  Proof. Consider a minimal drawing of G, and let p be a number between 0 and 1 (to be chosen later). Now we generate a subgraph of G, by selecting the vertices of G to lie in the subgraph with probability p, independently from each other. The induced subgraph that we obtain that way will be called Gp . Let np , mp , Xp be the random variables counting the number of vertices, of edges, and of crossings in Gp . Since cr(G) − m + 3n ≥ 0 holds by (4) for any graph, we certainly have E(Xp − mp + 3np ) ≥ 0. Now we proceed to compute the individual expectations E(np ), E(mp ) and E(Xp ). Clearly, E(np ) = pn and E(mp ) = p2 m, since an edge appears in Gp if and only if both its endvertices do. And finally, E(Xp ) = p4 cr(G), since a crossing is present in Gp if and only if all four (distinct!) vertices involved are there.

268

Probability makes counting (sometimes) easy By linearity of expectation we thus find 0 ≤ E(Xp ) − E(mp ) + 3E(np ) = p4 cr(G) − p2 m + 3pn, which is cr(G) ≥

p2 m − 3pn m 3n = 2 − 3. 4 p p p

(5)

Here comes the punch line: Set p := 4n m (which is at most 1 by our assumption), then (5) becomes @ A 1 4m 1 m3 3n cr(G) ≥ = − , 64 (n/m)2 (n/m)3 64 n2 and this is it.



Paul Erd˝os would have loved to see this proof.

References [1] M. A JTAI , V. C HVÁTAL , M. N EWBORN & E. S ZEMERÉDI : Crossing-free subgraphs, Annals of Discrete Math. 12 (1982), 9-12. [2] N. A LON & J. S PENCER : The Probabilistic Method, Third edition, WileyInterscience 2008. ˝ : Some remarks on the theory of graphs, Bulletin Amer. Math. Soc. [3] P. E RD OS 53 (1947), 292-294. ˝ : Graph theory and probability, Canadian J. Math. 11 (1959), 34-38. [4] P. E RD OS ˝ : On a combinatorial problem I, Nordisk Math. Tidskrift 11 (1963), [5] P. E RD OS 5-10. ˝ & R. K. G UY: Crossing number problems, Amer. Math. Monthly [6] P. E RD OS 80 (1973), 52-58. ˝ & A. R ÉNYI : On the evolution of random graphs, Magyar Tud. [7] P. E RD OS Akad. Mat. Kut. Int. Közl. 5 (1960), 17-61.

[8] T. L EIGHTON : Complexity Issues in VLSI, MIT Press, Cambridge MA 1983. [9] L. A. S ZÉKELY: Crossing numbers and hard Erd˝os problems in discrete geometry, Combinatorics, Probability, and Computing 6 (1997), 353-358.

Proofs from THE BOOK

M. Aigner, G.M. Ziegler, Proofs from THE BOOK, DOI 10.1007/978-3-642-00856-6, © Springer-Verlag Berlin Heidelberg 2013

269

About the Illustrations

We are happy to have the possibility and privilege to illustrate this volume with wonderful original drawings by Karl Heinrich Hofmann (Darmstadt). Thank you! The regular polyhedra on page 76 and the fold-out map of a flexible sphere on page 84 are by WAF Ruppert (Vienna). Jürgen Richter-Gebert (Munich) provided the two illustrations on page 78, and Ronald Wotzlaw wrote the nice postscript graphics for page 134. Page 231 features the Weisman Art Museum in Minneapolis designed by Frank Gehry. The photo of its west façade is by Chris Faust. The floorplan is of the Dolly Fiterman Riverview Gallery behind the west façade. The portraits of Bertrand, Cantor, Erd˝os, Euler, Fermat, Herglotz, Hermite, Hilbert, Pólya, Littlewood, and Sylvester are all from the photo archives of the Mathematisches Forschungsinstitut Oberwolfach, with permission. (Many thanks to Annette Disch!) The Gauss portrait on page 23 is a lithograph by Siegfried Detlev Bendixen published in Astronomische Nachrichten 1828, as provided by Wikipedia. The picture of Hermite is from the first volume of his collected works. The Eisenstein portrait is reproduced with friendly permission by Prof. Karin Reich from a collection of portrait cards owned by the Mathematische Gesellschaft Hamburg. The portrait stamps of Buffon, Chebyshev, Euler, and Ramanujan are from Jeff Miller’s mathematical stamps website http://jeff560.tripod.com with his generous permission. The photo of Claude Shannon was provided by the MIT Museum and is here reproduced with their permission. The portrait of Cayley is taken from the “Photoalbum für Weierstraß” (edited by Reinhard Bölling, Vieweg 1994), with permission from the Kunstbibliothek, Staatliche Museen zu Berlin, Preussischer Kulturbesitz. The Cauchy portrait is reproduced with permission from the Collections de l’École Polytechnique, Paris. The picture of Fermat is reproduced from Stefan Hildebrandt and Anthony Tromba: The Parsimonious Universe. Shape and Form in the Natural World, Springer-Verlag, New York 1996. The portrait of Ernst Witt is from volume 426 (1992) of the Journal für die Reine und Angewandte Mathematik, with permission by Walter de Gruyter Publishers. It was taken around 1941. The photo of Karol Borsuk was taken in 1967 by Isaac Namioka, and is reproduced with his kind permission. We thank Dr. Peter Sperner (Braunschweig) for the portrait of his father, and Vera Sós for the photo of Paul Turán. Thanks to Noga Alon for the portrait of A. Nilli!

Index

acyclic directed graph, 196 addition theorems, 150 adjacency matrix, 248 adjacent vertices, 66 antichain, 179 arithmetic mean, 119 art gallery theorem, 232 average degree, 76 average number of divisors, 164 Bernoulli numbers, 48, 152 Bertrand’s postulate, 7 bijection, 103, 207 Binet–Cauchy formula, 197, 203 binomial coefficient, 13 bipartite graph, 67, 223 birthday paradox, 185 Bolyai–Gerwien Theorem, 53 Borsuk’s conjecture, 95 Borsuk–Ulam theorem, 252 Bricard’s condition, 57 Brouwer’s fixed point theorem, 169 Buffon’s needle problem, 155 Calkin–Wilf tree, 105 Cantor–Bernstein theorem, 110 capacity, 242 cardinal number, 103 cardinality, 103, 115 Cauchy’s arm lemma, 82 Cauchy’s minimum principle, 127 Cauchy’s rigidity theorem, 81 Cauchy–Schwarz inequality, 119 Cayley’s formula, 201 center, 31 centralizer, 31 centrally symmetric, 61 chain, 179 channel, 241 Chebyshev polynomials, 144

Chebyshev’s theorem, 140 chromatic number, 221, 251 class formula, 32 clique, 67, 235, 242 clique number, 237 2-colorable set system, 261 combinatorially equivalent, 61 comparison of coefficients, 47 complete bipartite graph, 66 complete graph, 66 complex polynomial, 139 components of a graph, 67 cone lemma, 56 confusion graph, 241 congruent, 61 connected, 67 connected components, 67 continuum, 109 continuum hypothesis, 112 convex polytope, 59 convex vertex, 232 cosine polynomial, 143 countable, 103 coupon collector’s problem, 186 critical family, 182 crossing lemma, 267 crossing number, 266 cube, 60 cycle, 67 C4 -condition, 257 C4 -free graph, 166 degree, 76 dense, 114 determinants, 195 dihedral angle, 57 dimension, 109 dimension of a graph, 162 Dinitz problem, 221 directed graph, 223

272

Index division ring, 31 double counting, 164 dual graph, 75, 227 edge of a graph, 66 edge of a polyhedron, 60 elementary polygon, 79 equal size, 103 equicomplementability, 54 equicomplementable polyhedra, 53 equidecomposability, 54 equidecomposable polyhedra, 53 Erd˝os–Ko–Rado theorem, 180 Euler’s criterion, 24 Euler’s function, 28 Euler’s polyhedron formula, 75 Euler’s series, 43 even function, 152 expectation, 94 face, 60, 75 facet, 60 Fermat number, 3 finite field, 31 finite fields, 28 finite set system, 179 forest, 67 formal power series, 207 four-color theorem, 227 friendship theorem, 257 fundamental theorem of algebra, 127

hyper-binary representation, 106 incidence matrix, 64, 164 incident, 66 indegree, 223 independence number, 241, 251, 264 independent set, 67, 221 induced subgraph, 67, 222 inequalities, 119 infinite products, 207 initial ordinal number, 116 intersecting family, 180, 251 involution, 20 irrational numbers, 35 isomorphic graphs, 67 Jacobi determinants, 45 kernel, 223 Kneser graph, 251 Kneser’s conjecture, 252

Gale’s theorem, 253 Gauss lemma, 25 Gauss sum, 27 general position, 253 geometric mean, 119 Gessel–Viennot lemma, 195 girth, 263 golden section, 245 graph, 66 graph coloring, 227 graph of a polytope, 60

labeled tree, 201 Lagrange’s theorem, 4 Latin rectangle, 214 Latin square, 213, 221 lattice, 79 lattice basis, 80 lattice paths, 195 lattice points, 26 law of quadratic reciprocity, 24 Legendre symbol, 23 Legendre’s theorem, 8 lexicographically smallest solution, 56 line graph, 226 linear extension, 176 linearity of expectation, 94, 156 list chromatic number, 222 list coloring, 222, 228 Littlewood–Offord problem, 145 loop, 66 Lovász’ theorem, 247 Lovász umbrella, 244 Lyusternik–Shnirel’man theorem, 252

harmonic mean, 119 harmonic number, 10 Herglotz trick, 149 Hilbert’s third problem, 53

Markov’s inequality, 94 marriage theorem, 182 matching, 224 matrix of rank 1, 96

Index matrix-tree theorem, 203 Mersenne number, 4 Minkowski symmetrization, 91 mirror image, 61 monotone subsequences, 162 Monsky’s Theorem, 133 multiple edges, 66 museum guards, 231 Mycielski graph, 264 near-triangulated plane graph, 228 nearly-orthogonal vectors, 96 needles, 155 neighbors, 66 Newman’s function, 107 non-Archimedean real valuation, 132 non-Archimedean valuation, 136 obtuse angle, 89 odd function, 150 order of a group element, 4 ordered abelian group, 136 ordered set, 115 ordinal number, 115 orthonormal representation, 244 outdegree, 223 p-adic value, 132 partial Latin square, 213 partition, 207 partition identities, 207 path, 67 path matrix, 195 pearl lemma, 55 pentagonal numbers, 209 periodic function, 150 Petersen graph, 251 Pick’s theorem, 79 pigeon-hole principle, 161 planar graph, 75 plane graph, 75, 228 point configuration, 69 polygon, 59 polyhedron, 53, 59 polynomial with real roots, 121, 142 polytope, 89 prime field, 18 prime number, 3, 7

273 prime number theorem, 10 probabilistic method, 261 probability distribution, 236 probability space, 94 product of graphs, 241 projective plane, 167 quadratic nonresidue, 23 quadratic reciprocity, 24 quadratic residue, 23 rainbow triangle, 133 Ramsey number, 262 random variable, 94 rate of transmission, 241 red-blue segment, 135 refining sequence, 205 Riemann zeta function, 49 riffle shuffles, 191 Rogers–Ramanujan identities, 211 rooted forest, 205 roots of unity, 33 scalar product, 96 Schönhardt’s polyhedron, 232 segment, 54 Shannon capacity, 242 shuffling cards, 185 simple graph, 66 simplex, 60 size of a set, 103 slope problem, 69 speed of convergence, 47 Sperner’s lemma, 169 Sperner’s theorem, 179 squares, 18 stable matching, 224 star, 65 Stern’s diatomic series, 104 Stirling’s formula, 11 stopping rules, 188 subgraph, 67 sums of two squares, 17 Sylvester’s theorem, 13 Sylvester–Gallai theorem, 63, 78 system of distinct representatives, 181

274

Index tangential rectangle, 121 tangential triangle, 121 top-in-at-random shuffles, 187 touching simplices, 85 tree, 67 triangle-free graph, 264 Turán graph, 235 Turán’s graph theorem, 235 two square theorem, 17 umbrella, 244 unimodal, 12 unit d-cube, 60

valuation ring, 136 valuations, 131, 136 vertex, 60, 66 vertex degree, 76, 165, 222 vertex-disjoint path system, 195 volume, 84 weighted directed graph, 195 well-ordered, 115 well-ordering theorem, 115 windmill graph, 257 zero-error capacity, 242 Zorn’s lemma, 137